Don't try and install kompute artifacts.

stop using so many static objects so we can tear down and bring up vulkan
on new devices in the same runtime.

.gitignore

@@ -75,4 +75,3 @@ tests/test-quantize-fns
 tests/test-quantize-perf
 tests/test-sampling
 tests/test-tokenizer-0
-

CMakeLists.txt

@@ -67,13 +67,16 @@ endif()
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
-option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                OFF)
+option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
+#option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
+set(LLAMA_CUDA_MMQ_Y       "64" CACHE STRING "llama: y tile size for mmq CUDA kernels")
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
-option(LLAMA_CUDA_DMMV_F16                   "llama: use 16 bit floats for dmmv CUDA kernels"   OFF)
+option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
+option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
@@ -251,6 +254,10 @@ if (LLAMA_CUBLAS)
         set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+#        if (LLAMA_CUDA_CUBLAS)
+#            add_compile_definitions(GGML_CUDA_CUBLAS)
+#        endif()
+        add_compile_definitions(GGML_CUDA_MMQ_Y=${LLAMA_CUDA_MMQ_Y})
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -259,8 +266,8 @@ if (LLAMA_CUBLAS)
         if (DEFINED LLAMA_CUDA_DMMV_Y)
             add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_DMMV_Y}) # for backwards compatibility
         endif()
-        if (LLAMA_CUDA_DMMV_F16)
-            add_compile_definitions(GGML_CUDA_DMMV_F16)
+        if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
+            add_compile_definitions(GGML_CUDA_F16)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
@@ -271,10 +278,14 @@ if (LLAMA_CUBLAS)
         endif()
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        if (LLAMA_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "60;61") # needed for f16 CUDA intrinsics
+        # 52 == lowest CUDA 12 standard
+        # 60 == f16 CUDA intrinsics
+        # 61 == integer CUDA intrinsics
+        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
+        if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "52;61") # lowest CUDA 12 standard + lowest for integer intrinsics
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -346,6 +357,126 @@ if (LLAMA_CLBLAST)
     endif()
 endif()
 
+if (LLAMA_KOMPUTE)
+    find_package(Vulkan COMPONENTS glslc REQUIRED)
+    find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
+    if (NOT glslc_executable)
+        message(FATAL_ERROR "glslc not found")
+    endif()
+
+    function(compile_shader)
+      set(options)
+      set(oneValueArgs)
+      set(multiValueArgs SOURCES)
+      cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+      foreach(source ${compile_shader_SOURCES})
+        set(spv_file ${source}.spv)
+        add_custom_command(
+            OUTPUT ${spv_file}
+            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMENT "Compiling ${source} to ${source}.spv"
+        )
+
+        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+        set(FILE_NAME "shader${RAW_FILE_NAME}")
+        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+        add_custom_command(
+          OUTPUT ${OUTPUT_HEADER_FILE}
+          COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+          COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+          COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+          COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+          COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+          COMMAND xxd -i ${spv_file} >> ${OUTPUT_HEADER_FILE}
+          COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+          COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+          DEPENDS ${spv_file}
+          COMMENT "Converting to hpp: ${FILE_NAME}"
+        )
+      endforeach()
+    endfunction()
+
+    if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
+        message(STATUS "Kompute found")
+        add_subdirectory(kompute)
+
+        # Compile our shaders
+        compile_shader(SOURCES
+          kompute/op_scale.comp
+          kompute/op_add.comp
+          kompute/op_addrow.comp
+          kompute/op_mul.comp
+          kompute/op_mulrow.comp
+          kompute/op_silu.comp
+          kompute/op_relu.comp
+          kompute/op_gelu.comp
+          kompute/op_softmax.comp
+          kompute/op_norm.comp
+          kompute/op_rmsnorm.comp
+          kompute/op_diagmask.comp
+          kompute/op_mul_mat_f16.comp
+          kompute/op_mul_mat_q4_0.comp
+          kompute/op_mul_mat_q4_1.comp
+          kompute/op_getrows_f16.comp
+          kompute/op_getrows_q4_0.comp
+          kompute/op_getrows_q4_1.comp
+          kompute/op_rope.comp
+          kompute/op_cpy_f16_f16.comp
+          kompute/op_cpy_f16_f32.comp
+          kompute/op_cpy_f32_f16.comp
+          kompute/op_cpy_f32_f32.comp
+        )
+
+        # Create a custom target for our generated shaders
+        add_custom_target(generated_shaders DEPENDS
+          shaderop_scale.h
+          shaderop_add.h
+          shaderop_addrow.h
+          shaderop_mul.h
+          shaderop_mulrow.h
+          shaderop_silu.h
+          shaderop_relu.h
+          shaderop_gelu.h
+          shaderop_softmax.h
+          shaderop_norm.h
+          shaderop_rmsnorm.h
+          shaderop_diagmask.h
+          shaderop_mul_mat_f16.h
+          shaderop_mul_mat_q4_0.h
+          shaderop_mul_mat_q4_1.h
+          shaderop_getrows_f16.h
+          shaderop_getrows_q4_0.h
+          shaderop_getrows_q4_1.h
+          shaderop_rope.h
+          shaderop_cpy_f16_f16.h
+          shaderop_cpy_f16_f32.h
+          shaderop_cpy_f32_f16.h
+          shaderop_cpy_f32_f32.h
+        )
+
+        # Create a custom command that depends on the generated_shaders
+        add_custom_command(
+            OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan.stamp
+            COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan.stamp
+            DEPENDS generated_shaders
+            COMMENT "Ensuring shaders are generated before compiling ggml-vulkan.cpp"
+        )
+
+        # Add the stamp to the main sources to ensure dependency tracking
+        set(GGML_SOURCES_KOMPUTE ggml-vulkan.cpp ggml-vulkan.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan.stamp)
+        add_compile_definitions(GGML_USE_KOMPUTE)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
+    else()
+        message(WARNING "Kompute not found")
+    endif()
+endif()
+
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
         set(c_flags
@@ -357,6 +488,7 @@ if (LLAMA_ALL_WARNINGS)
             -Wshadow
             -Wstrict-prototypes
             -Wpointer-arith
+            -Wmissing-prototypes
         )
         set(cxx_flags
             -Wall
@@ -496,8 +628,11 @@ endif()
 add_library(ggml OBJECT
             ggml.c
             ggml.h
+            ggml-alloc.c
+            ggml-alloc.h
             ${GGML_SOURCES_CUDA}
             ${GGML_SOURCES_OPENCL}
+            ${GGML_SOURCES_KOMPUTE}
             ${GGML_SOURCES_METAL}
             ${GGML_SOURCES_MPI}
             ${GGML_SOURCES_EXTRA}

LICENSE_SOM.txt

@@ -0,0 +1,30 @@
+Software for Open Models License (SOM)
+Version 1.0 dated August 30th, 2023
+
+This license governs use of the accompanying Software. If you use the Software, you accept this license. If you do not accept the license, do not use the Software.
+
+This license is intended to encourage open release of models created, modified, processed, or otherwise used via the Software under open licensing terms, and should be interpreted in light of that intent.
+
+1. Definitions
+The “Licensor” is the person or entity who is making the Software available under this license. “Software” is the software made available by Licensor under this license.
+A “Model” is the output of a machine learning algorithm, and excludes the Software.
+“Model Source Materials” must include the Model and model weights, and may include any input data, input data descriptions, documentation or training descriptions for the Model.
+“Open Licensing Terms” means: (a) any open source license approved by the Open Source Initiative, or (b) any other terms that make the Model Source Materials publicly available free of charge, and allow recipients to use, modify and distribute the Model Source Materials. Terms described in (b) may include reasonable restrictions such as non-commercial or non-production limitations, or require use in compliance with law.
+
+2. Grant of Rights. Subject to the conditions and limitations in section 3:
+(A) Copyright Grant. Licensor grants you a non-exclusive, worldwide, royalty-free copyright license to copy, modify, and distribute the Software and any modifications of the Software you create under this license. The foregoing license includes without limitation the right to create, modify, and use Models using this Software.
+
+(B) Patent Grant. Licensor grants you a non-exclusive, worldwide, royalty-free license, under any patents owned or controlled by Licensor, to make, have made, use, sell, offer for sale, import, or otherwise exploit the Software.  No license is granted to patent rights that are not embodied in the operation of the Software in the form provided by Licensor.
+
+3. Conditions and Limitations
+(A) Model Licensing and Access. If you use the Software to create, modify, process, or otherwise use any Model, including usage to create inferences with a Model, whether or not you make the Model available to others, you must make that Model Source Materials publicly available under Open Licensing Terms. 
+
+(B) No Re-Licensing. If you redistribute the Software, or modifications to the Software made under the license granted above, you must make it available only under the terms of this license. You may offer additional terms such as warranties, maintenance and support, but You, and not Licensor, are responsible for performing such terms.
+
+(C) No Trademark License. This license does not grant you rights to use the Licensor’s name, logo, or trademarks.
+
+(D) If you assert in writing a claim against any person or entity alleging that the use of the Software infringes any patent, all of your licenses to the Software under Section 2 end automatically as of the date you asserted the claim.
+
+(E) If you distribute any portion of the Software, you must retain all copyright, patent, trademark, and attribution notices that are present in the Software, and you must include a copy of this license.
+
+(F) The Software is licensed “as-is.” You bear the entire risk of using it. Licensor gives You no express warranties, guarantees or conditions. You may have additional consumer rights under your local laws that this license cannot change. To the extent permitted under your local laws, the Licensor disclaims and excludes the implied warranties of merchantability, fitness for a particular purpose and non-infringement. To the extent this disclaimer is unlawful, you, and not Licensor, are responsible for any liability.

Makefile

@@ -63,7 +63,8 @@ ifdef LLAMA_SERVER_VERBOSE
 endif
 
 # warnings
-CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith
+CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \
+			-Wmissing-prototypes
 CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
 
 # OS specific
@@ -141,6 +142,28 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	#CXXFLAGS += -mssse3
 endif
 
+ifneq ($(filter aarch64%,$(UNAME_M)),)
+	# Apple M1, M2, etc.
+	# Raspberry Pi 3, 4, Zero 2 (64-bit)
+	CFLAGS   += -mcpu=native
+	CXXFLAGS += -mcpu=native
+endif
+
+ifneq ($(filter armv6%,$(UNAME_M)),)
+	# Raspberry Pi 1, Zero
+	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+endif
+
+ifneq ($(filter armv7%,$(UNAME_M)),)
+	# Raspberry Pi 2
+	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+endif
+
+ifneq ($(filter armv8%,$(UNAME_M)),)
+	# Raspberry Pi 3, 4, Zero 2 (32-bit)
+	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
+endif
+
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
@@ -193,7 +216,7 @@ ifdef LLAMA_CUBLAS
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
-	NVCCFLAGS = --forward-unknown-to-host-compiler
+	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else
@@ -219,19 +242,30 @@ else ifdef LLAMA_CUDA_DMMV_Y
 else
 	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
+ifdef LLAMA_CUDA_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
+endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	NVCCFLAGS += -DGGML_CUDA_DMMV_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+ifdef LLAMA_CUDA_MMQ_Y
+	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=$(LLAMA_CUDA_MMQ_Y)
+else
+	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=64
+endif # LLAMA_CUDA_MMQ_Y
+#ifdef LLAMA_CUDA_CUBLAS
+#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
 	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
@@ -258,28 +292,6 @@ ifdef LLAMA_METAL
 	OBJS     += ggml-metal.o
 endif # LLAMA_METAL
 
-ifneq ($(filter aarch64%,$(UNAME_M)),)
-	# Apple M1, M2, etc.
-	# Raspberry Pi 3, 4, Zero 2 (64-bit)
-	CFLAGS   += -mcpu=native
-	CXXFLAGS += -mcpu=native
-endif
-
-ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, Zero
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
-endif
-
-ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 2
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
-endif
-
-ifneq ($(filter armv8%,$(UNAME_M)),)
-	# Raspberry Pi 3, 4, Zero 2 (32-bit)
-	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
-endif
-
 ifdef LLAMA_METAL
 ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
@@ -317,12 +329,20 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
+ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+OBJS += ggml-alloc.o
+
+llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 common.o: examples/common.cpp examples/common.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+console.o: examples/console.cpp examples/console.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 grammar-parser.o: examples/grammar-parser.cpp examples/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
@@ -336,7 +356,7 @@ clean:
 # Examples
 #
 
-main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
@@ -394,13 +414,13 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-double-float: tests/test-double-float.c build-info.h ggml.o llama.o common.o $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.c build-info.h ggml.o llama.o common.o $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
 
-tests/test-opt: tests/test-opt.c build-info.h ggml.o llama.o common.o $(OBJS)
+tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
 
 tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS)

README.md

@@ -77,9 +77,10 @@ as the main playground for developing new features for the [ggml](https://github
 **Supported models:**
 
 - [X] LLaMA 🦙
+- [x] LLaMA 2 🦙🦙
 - [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
 - [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
-- [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca)
+- [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
 - [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
@@ -87,6 +88,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 - [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
+- [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
 
 **Bindings:**
 
@@ -399,12 +401,16 @@ Building the program with BLAS support may lead to some performance improvements
 
   The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
 
+<!---
+  | LLAMA_CUDA_CUBLAS       | Boolean                |   false | Use cuBLAS instead of custom CUDA kernels for prompt processing. Faster for all quantization formats except for q4_0 and q8_0, especially for k-quants. Increases VRAM usage (700 MiB for 7b, 970 MiB for 13b, 1430 MiB for 33b). |
+--->
   | Option                  | Legal values           | Default | Description |
   |-------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_MMQ_Y        | Positive integer >= 32 |      64 | Tile size in y direction when using the custom CUDA kernels for prompt processing. Higher values can be faster depending on the amount of shared memory available. Power of 2 heavily recommended. |
   | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
   | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
 
 - #### CLBlast
@@ -487,6 +493,9 @@ Building the program with BLAS support may lead to some performance improvements
 # obtain the original LLaMA model weights and place them in ./models
 ls ./models
 65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
+  # [Optional] for models using BPE tokenizers
+  ls ./models
+  65B 30B 13B 7B vocab.json
 
 # install Python dependencies
 python3 -m pip install -r requirements.txt
@@ -494,6 +503,9 @@ python3 -m pip install -r requirements.txt
 # convert the 7B model to ggml FP16 format
 python3 convert.py models/7B/
 
+  # [Optional] for models using BPE tokenizers
+  python convert.py models/7B/ --vocabtype bpe
+
 # quantize the model to 4-bits (using q4_0 method)
 ./quantize ./models/7B/ggml-model-f16.bin ./models/7B/ggml-model-q4_0.bin q4_0
 
@@ -650,6 +662,19 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
 - The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
 - Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
 
+### Obtaining and using the Facebook LLaMA 2 model
+
+- Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
+- Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
+  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGML)
+  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGML)
+  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGML)
+  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML)
+  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML)
+  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML)
+- Specify `-eps 1e-5` for best generation quality
+- Specify `-gqa 8` for 70B models to work
+
 ### Verifying the model files
 
 Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.

build.zig

@@ -1,68 +1,87 @@
+// Compatible with Zig Version 0.11.0
 const std = @import("std");
-const commit_hash = @embedFile(".git/refs/heads/master");
+const Compile = std.Build.Step.Compile;
+const ConfigHeader = std.Build.Step.ConfigHeader;
+const Mode = std.builtin.Mode;
+const CrossTarget = std.zig.CrossTarget;
+
+const Maker = struct {
+    builder: *std.build.Builder,
+    target: CrossTarget,
+    optimize: Mode,
+    config_header: *ConfigHeader,
+
+    const cflags = .{"-std=c11"};
+    const cxxflags = .{"-std=c++11"};
+
+    fn init(builder: *std.build.Builder) Maker {
+        const commit_hash = @embedFile(".git/refs/heads/master");
+        const config_header = builder.addConfigHeader(
+            .{ .style = .blank, .include_path = "build-info.h" },
+            .{
+                .BUILD_NUMBER = 0,
+                .BUILD_COMMIT = commit_hash[0 .. commit_hash.len - 1], // omit newline
+            },
+        );
+        return Maker{
+            .builder = builder,
+            .target = builder.standardTargetOptions(.{}),
+            .optimize = builder.standardOptimizeOption(.{}),
+            .config_header = config_header,
+        };
+    }
+
+    fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
+        const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
+        if (std.mem.endsWith(u8, src, ".c")) {
+            o.addCSourceFiles(&.{src}, &cflags);
+            o.linkLibC();
+        } else {
+            o.addCSourceFiles(&.{src}, &cxxflags);
+            o.linkLibCpp();
+        }
+        o.addIncludePath(.{ .path = "." });
+        o.addIncludePath(.{ .path = "./examples" });
+        return o;
+    }
+
+    fn exe(m: *const Maker, name: []const u8, src: []const u8, deps: []const *Compile) *Compile {
+        const e = m.builder.addExecutable(.{ .name = name, .target = m.target, .optimize = m.optimize });
+        e.addIncludePath(.{ .path = "." });
+        e.addIncludePath(.{ .path = "./examples" });
+        e.addCSourceFiles(&.{src}, &cxxflags);
+        for (deps) |d| e.addObject(d);
+        e.linkLibC();
+        e.linkLibCpp();
+        e.addConfigHeader(m.config_header);
+        m.builder.installArtifact(e);
+
+        // Currently a bug is preventing correct linking for optimized builds for Windows:
+        // https://github.com/ziglang/zig/issues/15958
+        if (e.target.isWindows()) {
+            e.want_lto = false;
+        }
+        return e;
+    }
+};
 
-// Zig Version: 0.11.0-dev.3986+e05c242cd
 pub fn build(b: *std.build.Builder) void {
-    const target = b.standardTargetOptions(.{});
-    const optimize = b.standardOptimizeOption(.{});
+    const make = Maker.init(b);
 
-    const config_header = b.addConfigHeader(
-        .{ .style = .blank, .include_path = "build-info.h" },
-        .{
-            .BUILD_NUMBER = 0,
-            .BUILD_COMMIT = commit_hash[0 .. commit_hash.len - 1], // omit newline
-        },
-    );
+    const ggml = make.obj("ggml", "ggml.c");
+    const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
+    const llama = make.obj("llama", "llama.cpp");
+    const common = make.obj("common", "examples/common.cpp");
+    const grammar_parser = make.obj("grammar-parser", "examples/grammar-parser.cpp");
 
-    const lib = b.addStaticLibrary(.{
-        .name = "llama",
-        .target = target,
-        .optimize = optimize,
-    });
-    lib.linkLibC();
-    lib.linkLibCpp();
-    lib.addIncludePath(".");
-    lib.addIncludePath("./examples");
-    lib.addConfigHeader(config_header);
-    lib.addCSourceFiles(&.{"ggml.c"}, &.{"-std=c11"});
-    lib.addCSourceFiles(&.{"llama.cpp"}, &.{"-std=c++11"});
-    b.installArtifact(lib);
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, llama, common, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, llama, common });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, llama, common });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, llama });
 
-    const examples = .{
-        "main",
-        "baby-llama",
-        "embedding",
-        "metal",
-        "perplexity",
-        "quantize",
-        "quantize-stats",
-        "save-load-state",
-        "server",
-        "simple",
-        "train-text-from-scratch",
-    };
-
-    inline for (examples) |example_name| {
-        const exe = b.addExecutable(.{
-            .name = example_name,
-            .target = target,
-            .optimize = optimize,
-        });
-        exe.addIncludePath(".");
-        exe.addIncludePath("./examples");
-        exe.addConfigHeader(config_header);
-        exe.addCSourceFiles(&.{
-            std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{ example_name, example_name }),
-            "examples/common.cpp",
-        }, &.{"-std=c++11"});
-        exe.linkLibrary(lib);
-        b.installArtifact(exe);
-
-        const run_cmd = b.addRunArtifact(exe);
-        run_cmd.step.dependOn(b.getInstallStep());
-        if (b.args) |args| run_cmd.addArgs(args);
-
-        const run_step = b.step("run-" ++ example_name, "Run the app");
-        run_step.dependOn(&run_cmd.step);
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, llama, common, grammar_parser });
+    if (server.target.isWindows()) {
+        server.linkSystemLibrary("ws2_32");
     }
 }

convert.py

@@ -133,7 +133,7 @@ TENSORS_SET = set(TENSORS_LIST)
 
 def find_n_mult(n_ff: int, n_embd: int) -> int:
     # hardcoded magic range
-    for n_mult in range(256, 1, -1):
+    for n_mult in range(8192, 1, -1):
         calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
         if calc_ff == n_ff:
             return n_mult
@@ -141,11 +141,12 @@ def find_n_mult(n_ff: int, n_embd: int) -> int:
 
 @dataclass
 class Params:
-    n_vocab: int
-    n_embd:  int
-    n_mult:  int
-    n_head:  int
-    n_layer: int
+    n_vocab:   int
+    n_embd:    int
+    n_mult:    int
+    n_head:    int
+    n_layer:   int
+    n_kv_head: Optional[int]  # This parameter is only used for Llama 2
 
     @staticmethod
     def guessed(model: 'LazyModel') -> 'Params':
@@ -167,11 +168,12 @@ class Params:
         n_head=n_embd // 128 # guessed
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = 256,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = 256,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = None,
         )
 
     @staticmethod
@@ -183,15 +185,17 @@ class Params:
         n_head  = config["num_attention_heads"];
         n_layer = config["num_hidden_layers"];
         n_ff    = config["intermediate_size"];
+        n_kv_head = config.get("num_key_value_heads")
 
         n_mult = find_n_mult(n_ff, n_embd);
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = n_mult,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = n_mult,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = n_kv_head,
         )
 
     # LLaMA v2 70B params.json
@@ -200,21 +204,22 @@ class Params:
     def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
         config = json.load(open(config_path))
 
-        n_vocab = config["vocab_size"];
-        n_embd  = config["dim"];
-        n_head  = config["n_heads"];
-        n_layer = config["n_layers"];
-        n_mult  = config["multiple_of"];
+        n_vocab   = config["vocab_size"];
+        n_embd    = config["dim"];
+        n_head    = config["n_heads"];
+        n_layer   = config["n_layers"];
+        n_mult    = config["multiple_of"];
 
         if n_vocab == -1:
             n_vocab = model["tok_embeddings.weight"].shape[0]
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = n_mult,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = n_mult,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = None,
         )
 
     @staticmethod
@@ -317,10 +322,12 @@ class GGMLVocab:
 Vocab = Union[SentencePieceVocab, GGMLVocab]
 
 
-def permute(weights: NDArray, n_head: int) -> NDArray:
+def permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] = None) -> NDArray:
+    if n_kv_head is not None and n_head != n_kv_head:
+        n_head //= n_kv_head
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-                   .swapaxes(1, 2)
-                   .reshape(weights.shape))
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
 
 
 def dequantize_q4(qvalues_pack32: NDArray, scales: NDArray, addends: Optional[NDArray], g_idx: Optional[NDArray]) -> NDArray:
@@ -368,7 +375,7 @@ class Tensor(metaclass=ABCMeta):
     @abstractmethod
     def astype(self, data_type: DataType) -> 'Tensor': ...
     @abstractmethod
-    def permute(self, n_head: int) -> 'Tensor': ...
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'Tensor': ...
     @abstractmethod
     def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
     @abstractmethod
@@ -406,8 +413,8 @@ class UnquantizedTensor(Tensor):
         r = self.ndarray.shape[0] // 3
         return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
-    def permute(self, n_head: int) -> 'UnquantizedTensor':
-        return UnquantizedTensor(permute(self.ndarray, n_head))
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'UnquantizedTensor':
+        return UnquantizedTensor(permute(self.ndarray, n_head, n_kv_head))
 
 
 def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray:
@@ -455,26 +462,34 @@ class GGMLQuantizedTensor(Tensor):
     def to_ggml(self) -> 'GGMLQuantizedTensor':
         return self
 
-    def permute(self, n_head: int) -> 'GGMLQuantizedTensor':
-        return GGMLQuantizedTensor(permute(self.ndarray, n_head), self.shape, self.data_type)
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'GGMLQuantizedTensor':
+        return GGMLQuantizedTensor(permute(self.ndarray, n_head, n_kv_head), self.shape, self.data_type)
 
+    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
+
+    def part(self, n_part: int) -> 'UnquantizedTensor':
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
 GGMLCompatibleTensor = Union[UnquantizedTensor, GGMLQuantizedTensor]
 
 
 class DeferredPermutedTensor(Tensor):
-    def __init__(self, base: Tensor, n_head: int) -> None:
+    def __init__(self, base: Tensor, n_head: int, n_kv_head: Optional[int] = None) -> None:
         self.base = base
         self.n_head = n_head
+        self.n_kv_head = n_kv_head
         self.data_type = self.base.data_type
 
     def astype(self, data_type: DataType) -> Tensor:
-        return self.base.astype(data_type).permute(self.n_head)
+        return self.base.astype(data_type).permute(self.n_head, self.n_kv_head)
 
     def to_ggml(self) -> GGMLCompatibleTensor:
-        return self.base.to_ggml().permute(self.n_head)
+        return self.base.to_ggml().permute(self.n_head, self.n_kv_head)
 
-    def permute(self, n_head: int) -> Tensor:
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
         raise Exception("shouldn't permute twice")
 
 
@@ -566,8 +581,8 @@ class GPTQForLLaMaQuantizedTensor(Tensor):
         ret.data_type = QuantizedDataType(groupsize=new_groupsize, have_addends=True, have_g_idx=False)
         return ret
 
-    def permute(self, n_head: int) -> Tensor:
-        return DeferredPermutedTensor(self, n_head)
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
+        return DeferredPermutedTensor(self, n_head, n_kv_head)
 
     def to_ggml(self) -> GGMLQuantizedTensor:
         # The output format looks like this:
@@ -698,10 +713,10 @@ def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus:
     return ModelPlus(model, paths, format, vocab)
 
 
-def permute_lazy(lazy_tensor: LazyTensor, n_head: int) -> LazyTensor:
+def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_kv_head: Optional[int] = None) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute(n_head)
-    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+        return lazy_tensor.load().permute(n_head, n_kv_head)
+    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_kv_head}) ' + lazy_tensor.description)
 
 def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
     def load() -> Tensor:
@@ -726,7 +741,7 @@ def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel:
     for i in itertools.count():
         if f"model.layers.{i}.self_attn.q_proj.weight" in model:
             out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head)
-            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head)
+            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_kv_head)
             out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
         elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
             out[f"layers.{i}.attention.wq.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head)

examples/CMakeLists.txt

@@ -13,6 +13,8 @@ set(TARGET common)
 add_library(${TARGET} OBJECT
     common.h
     common.cpp
+    console.h
+    console.cpp
     grammar-parser.h
     grammar-parser.cpp
     )

examples/common.cpp

@@ -25,7 +25,6 @@
 #else
 #include <sys/ioctl.h>
 #include <unistd.h>
-#include <wchar.h>
 #endif
 
 #if defined(_MSC_VER)
@@ -195,6 +194,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = 1.0f/std::stof(argv[i]);
         } else if (arg == "--memory-f32") {
             params.memory_f16 = false;
         } else if (arg == "--top-p") {
@@ -329,6 +334,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.instruct = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
+        } else if (arg == "--simple-io") {
+            params.simple_io = true;
         } else if (arg == "--color") {
             params.use_color = true;
         } else if (arg == "--mlock") {
@@ -352,7 +359,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #ifdef GGML_USE_CUBLAS
             params.main_gpu = std::stoi(argv[i]);
 #else
-      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
 #endif
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
@@ -376,13 +383,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 }
             }
 #else
-      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--mul-mat-q" || arg == "-mmq") {
+#ifdef GGML_USE_CUBLAS
+            params.mul_mat_q = true;
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--low-vram" || arg == "-lv") {
 #ifdef GGML_USE_CUBLAS
             params.low_vram = true;
 #else
-      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -402,8 +415,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.antiprompt.push_back(argv[i]);
         } else if (arg == "--perplexity") {
             params.perplexity = true;
-        } else if (arg == "--perplexity-lines") {
-            params.perplexity_lines = true;
+        } else if (arg == "--hellaswag") {
+            params.hellaswag = true;
+        } else if (arg == "--hellaswag-tasks") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.hellaswag_tasks = std::stoi(argv[i]);
         } else if (arg == "--ignore-eos") {
             params.logit_bias[llama_token_eos()] = -INFINITY;
         } else if (arg == "--no-penalize-nl") {
@@ -551,16 +570,18 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "  --cfg-negative-prompt PROMPT \n");
     fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
     fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    fprintf(stdout, "  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
+    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
     fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
     fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
     fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
     fprintf(stdout, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
     fprintf(stdout, "  --perplexity          compute perplexity over each ctx window of the prompt\n");
-    fprintf(stdout, "  --perplexity-lines    compute perplexity over each line of the prompt\n");
-    fprintf(stdout, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    fprintf(stdout, "  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
+    fprintf(stdout, "  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
+    fprintf(stdout, "  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
     fprintf(stdout, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     if (llama_mlock_supported()) {
         fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
@@ -578,10 +599,14 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
     fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
+    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
+    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
+    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
 #endif
     fprintf(stdout, "  --mtest               compute maximum memory usage\n");
     fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
     fprintf(stdout, "  --verbose-prompt      print prompt before generation\n");
+    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stdout, "  -m FNAME, --model FNAME\n");
@@ -630,6 +655,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     lparams.main_gpu        = params.main_gpu;
     lparams.tensor_split    = params.tensor_split;
     lparams.low_vram        = params.low_vram;
+    lparams.mul_mat_q       = params.mul_mat_q;
     lparams.seed            = params.seed;
     lparams.f16_kv          = params.memory_f16;
     lparams.use_mmap        = params.use_mmap;
@@ -673,376 +699,3 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
     return std::make_tuple(model, lctx);
 }
-
-void console_init(console_state & con_st) {
-#if defined(_WIN32)
-    // Windows-specific console initialization
-    DWORD dwMode = 0;
-    con_st.hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
-    if (con_st.hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(con_st.hConsole, &dwMode)) {
-        con_st.hConsole = GetStdHandle(STD_ERROR_HANDLE);
-        if (con_st.hConsole != INVALID_HANDLE_VALUE && (!GetConsoleMode(con_st.hConsole, &dwMode))) {
-            con_st.hConsole = NULL;
-        }
-    }
-    if (con_st.hConsole) {
-        // Enable ANSI colors on Windows 10+
-        if (con_st.use_color && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
-            SetConsoleMode(con_st.hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING);
-        }
-        // Set console output codepage to UTF8
-        SetConsoleOutputCP(CP_UTF8);
-    }
-    HANDLE hConIn = GetStdHandle(STD_INPUT_HANDLE);
-    if (hConIn != INVALID_HANDLE_VALUE && GetConsoleMode(hConIn, &dwMode)) {
-        // Set console input codepage to UTF16
-        _setmode(_fileno(stdin), _O_WTEXT);
-
-        // Turn off ICANON (ENABLE_LINE_INPUT) and ECHO (ENABLE_ECHO_INPUT)
-        dwMode &= ~(ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT);
-        SetConsoleMode(hConIn, dwMode);
-    }
-#else
-    // POSIX-specific console initialization
-    struct termios new_termios;
-    tcgetattr(STDIN_FILENO, &con_st.prev_state);
-    new_termios = con_st.prev_state;
-    new_termios.c_lflag &= ~(ICANON | ECHO);
-    new_termios.c_cc[VMIN] = 1;
-    new_termios.c_cc[VTIME] = 0;
-    tcsetattr(STDIN_FILENO, TCSANOW, &new_termios);
-
-    con_st.tty = fopen("/dev/tty", "w+");
-    if (con_st.tty != nullptr) {
-        con_st.out = con_st.tty;
-    }
-
-    setlocale(LC_ALL, "");
-#endif
-}
-
-void console_cleanup(console_state & con_st) {
-    // Reset console color
-    console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
-
-#if !defined(_WIN32)
-    if (con_st.tty != nullptr) {
-        con_st.out = stdout;
-        fclose(con_st.tty);
-        con_st.tty = nullptr;
-    }
-    // Restore the terminal settings on POSIX systems
-    tcsetattr(STDIN_FILENO, TCSANOW, &con_st.prev_state);
-#endif
-}
-
-/* Keep track of current color of output, and emit ANSI code if it changes. */
-void console_set_color(console_state & con_st, console_color_t color) {
-    if (con_st.use_color && con_st.color != color) {
-        fflush(stdout);
-        switch(color) {
-            case CONSOLE_COLOR_DEFAULT:
-                fprintf(con_st.out, ANSI_COLOR_RESET);
-                break;
-            case CONSOLE_COLOR_PROMPT:
-                fprintf(con_st.out, ANSI_COLOR_YELLOW);
-                break;
-            case CONSOLE_COLOR_USER_INPUT:
-                fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_GREEN);
-                break;
-            case CONSOLE_COLOR_ERROR:
-                fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_RED);
-                break;
-        }
-        con_st.color = color;
-        fflush(con_st.out);
-    }
-}
-
-char32_t getchar32() {
-#if defined(_WIN32)
-    HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
-    wchar_t high_surrogate = 0;
-
-    while (true) {
-        INPUT_RECORD record;
-        DWORD count;
-        if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
-            return WEOF;
-        }
-
-        if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
-            wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
-            if (wc == 0) {
-                continue;
-            }
-
-            if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
-                high_surrogate = wc;
-                continue;
-            } else if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
-                if (high_surrogate != 0) { // Check if we have a high surrogate
-                    return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
-                }
-            }
-
-            high_surrogate = 0; // Reset the high surrogate
-            return static_cast<char32_t>(wc);
-        }
-    }
-#else
-    wchar_t wc = getwchar();
-    if (static_cast<wint_t>(wc) == WEOF) {
-        return WEOF;
-    }
-
-#if WCHAR_MAX == 0xFFFF
-    if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
-        wchar_t low_surrogate = getwchar();
-        if ((low_surrogate >= 0xDC00) && (low_surrogate <= 0xDFFF)) { // Check if the next wchar is a low surrogate
-            return (static_cast<char32_t>(wc & 0x03FF) << 10) + (low_surrogate & 0x03FF) + 0x10000;
-        }
-    }
-    if ((wc >= 0xD800) && (wc <= 0xDFFF)) { // Invalid surrogate pair
-        return 0xFFFD; // Return the replacement character U+FFFD
-    }
-#endif
-
-    return static_cast<char32_t>(wc);
-#endif
-}
-
-void pop_cursor(console_state & con_st) {
-#if defined(_WIN32)
-    if (con_st.hConsole != NULL) {
-        CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
-        GetConsoleScreenBufferInfo(con_st.hConsole, &bufferInfo);
-
-        COORD newCursorPosition = bufferInfo.dwCursorPosition;
-        if (newCursorPosition.X == 0) {
-            newCursorPosition.X = bufferInfo.dwSize.X - 1;
-            newCursorPosition.Y -= 1;
-        } else {
-            newCursorPosition.X -= 1;
-        }
-
-        SetConsoleCursorPosition(con_st.hConsole, newCursorPosition);
-        return;
-    }
-#endif
-    putc('\b', con_st.out);
-}
-
-int estimateWidth(char32_t codepoint) {
-#if defined(_WIN32)
-    return 1;
-#else
-    return wcwidth(codepoint);
-#endif
-}
-
-int put_codepoint(console_state & con_st, const char* utf8_codepoint, size_t length, int expectedWidth) {
-#if defined(_WIN32)
-    CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
-    if (!GetConsoleScreenBufferInfo(con_st.hConsole, &bufferInfo)) {
-        // go with the default
-        return expectedWidth;
-    }
-    COORD initialPosition = bufferInfo.dwCursorPosition;
-    DWORD nNumberOfChars = length;
-    WriteConsole(con_st.hConsole, utf8_codepoint, nNumberOfChars, &nNumberOfChars, NULL);
-
-    CONSOLE_SCREEN_BUFFER_INFO newBufferInfo;
-    GetConsoleScreenBufferInfo(con_st.hConsole, &newBufferInfo);
-
-    // Figure out our real position if we're in the last column
-    if (utf8_codepoint[0] != 0x09 && initialPosition.X == newBufferInfo.dwSize.X - 1) {
-        DWORD nNumberOfChars;
-        WriteConsole(con_st.hConsole, &" \b", 2, &nNumberOfChars, NULL);
-        GetConsoleScreenBufferInfo(con_st.hConsole, &newBufferInfo);
-    }
-
-    int width = newBufferInfo.dwCursorPosition.X - initialPosition.X;
-    if (width < 0) {
-        width += newBufferInfo.dwSize.X;
-    }
-    return width;
-#else
-    // we can trust expectedWidth if we've got one
-    if (expectedWidth >= 0 || con_st.tty == nullptr) {
-        fwrite(utf8_codepoint, length, 1, con_st.out);
-        return expectedWidth;
-    }
-
-    fputs("\033[6n", con_st.tty); // Query cursor position
-    int x1, x2, y1, y2;
-    int results = 0;
-    results = fscanf(con_st.tty, "\033[%d;%dR", &y1, &x1);
-
-    fwrite(utf8_codepoint, length, 1, con_st.tty);
-
-    fputs("\033[6n", con_st.tty); // Query cursor position
-    results += fscanf(con_st.tty, "\033[%d;%dR", &y2, &x2);
-
-    if (results != 4) {
-        return expectedWidth;
-    }
-
-    int width = x2 - x1;
-    if (width < 0) {
-        // Calculate the width considering text wrapping
-        struct winsize w;
-        ioctl(STDOUT_FILENO, TIOCGWINSZ, &w);
-        width += w.ws_col;
-    }
-    return width;
-#endif
-}
-
-void replace_last(console_state & con_st, char ch) {
-#if defined(_WIN32)
-    pop_cursor(con_st);
-    put_codepoint(con_st, &ch, 1, 1);
-#else
-    fprintf(con_st.out, "\b%c", ch);
-#endif
-}
-
-void append_utf8(char32_t ch, std::string & out) {
-    if (ch <= 0x7F) {
-        out.push_back(static_cast<unsigned char>(ch));
-    } else if (ch <= 0x7FF) {
-        out.push_back(static_cast<unsigned char>(0xC0 | ((ch >> 6) & 0x1F)));
-        out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
-    } else if (ch <= 0xFFFF) {
-        out.push_back(static_cast<unsigned char>(0xE0 | ((ch >> 12) & 0x0F)));
-        out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
-        out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
-    } else if (ch <= 0x10FFFF) {
-        out.push_back(static_cast<unsigned char>(0xF0 | ((ch >> 18) & 0x07)));
-        out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 12) & 0x3F)));
-        out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
-        out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
-    } else {
-        // Invalid Unicode code point
-    }
-}
-
-// Helper function to remove the last UTF-8 character from a string
-void pop_back_utf8_char(std::string & line) {
-    if (line.empty()) {
-        return;
-    }
-
-    size_t pos = line.length() - 1;
-
-    // Find the start of the last UTF-8 character (checking up to 4 bytes back)
-    for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
-        if ((line[pos] & 0xC0) != 0x80) break; // Found the start of the character
-    }
-    line.erase(pos);
-}
-
-bool console_readline(console_state & con_st, std::string & line) {
-    console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
-    if (con_st.out != stdout) {
-        fflush(stdout);
-    }
-
-    line.clear();
-    std::vector<int> widths;
-    bool is_special_char = false;
-    bool end_of_stream = false;
-
-    char32_t input_char;
-    while (true) {
-        fflush(con_st.out); // Ensure all output is displayed before waiting for input
-        input_char = getchar32();
-
-        if (input_char == '\r' || input_char == '\n') {
-            break;
-        }
-
-        if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
-            end_of_stream = true;
-            break;
-        }
-
-        if (is_special_char) {
-            console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
-            replace_last(con_st, line.back());
-            is_special_char = false;
-        }
-
-        if (input_char == '\033') { // Escape sequence
-            char32_t code = getchar32();
-            if (code == '[' || code == 0x1B) {
-                // Discard the rest of the escape sequence
-                while ((code = getchar32()) != (char32_t) WEOF) {
-                    if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
-                        break;
-                    }
-                }
-            }
-        } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
-            if (!widths.empty()) {
-                int count;
-                do {
-                    count = widths.back();
-                    widths.pop_back();
-                    // Move cursor back, print space, and move cursor back again
-                    for (int i = 0; i < count; i++) {
-                        replace_last(con_st, ' ');
-                        pop_cursor(con_st);
-                    }
-                    pop_back_utf8_char(line);
-                } while (count == 0 && !widths.empty());
-            }
-        } else {
-            int offset = line.length();
-            append_utf8(input_char, line);
-            int width = put_codepoint(con_st, line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
-            if (width < 0) {
-                width = 0;
-            }
-            widths.push_back(width);
-        }
-
-        if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
-            console_set_color(con_st, CONSOLE_COLOR_PROMPT);
-            replace_last(con_st, line.back());
-            is_special_char = true;
-        }
-    }
-
-    bool has_more = con_st.multiline_input;
-    if (is_special_char) {
-        replace_last(con_st, ' ');
-        pop_cursor(con_st);
-
-        char last = line.back();
-        line.pop_back();
-        if (last == '\\') {
-            line += '\n';
-            fputc('\n', con_st.out);
-            has_more = !has_more;
-        } else {
-            // llama will just eat the single space, it won't act as a space
-            if (line.length() == 1 && line.back() == ' ') {
-                line.clear();
-                pop_cursor(con_st);
-            }
-            has_more = false;
-        }
-    } else {
-        if (end_of_stream) {
-            has_more = false;
-        } else {
-            line += '\n';
-            fputc('\n', con_st.out);
-        }
-    }
-
-    fflush(con_st.out);
-    return has_more;
-}

examples/common.h

@@ -11,11 +11,6 @@
 #include <unordered_map>
 #include <tuple>
 
-#if !defined (_WIN32)
-#include <stdio.h>
-#include <termios.h>
-#endif
-
 //
 // CLI argument parsing
 //
@@ -70,7 +65,11 @@ struct gpt_params {
     std::string lora_adapter = "";  // lora adapter path
     std::string lora_base    = "";  // base model path for the lora adapter
 
-    bool low_vram          = false;   // if true, reduce VRAM usage at the cost of performance
+    bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
+
+    bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
+    bool mul_mat_q         = false; // if true, use experimental mul_mat_q kernels
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
@@ -81,12 +80,12 @@ struct gpt_params {
     bool embedding         = false; // get only sentence embedding
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
+    bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool penalize_nl       = true;  // consider newlines as a repeatable token
     bool perplexity        = false; // compute perplexity over the prompt
-    bool perplexity_lines  = false; // compute perplexity over each line of the prompt
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
@@ -113,42 +112,3 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
-
-//
-// Console utils
-//
-
-#define ANSI_COLOR_RED     "\x1b[31m"
-#define ANSI_COLOR_GREEN   "\x1b[32m"
-#define ANSI_COLOR_YELLOW  "\x1b[33m"
-#define ANSI_COLOR_BLUE    "\x1b[34m"
-#define ANSI_COLOR_MAGENTA "\x1b[35m"
-#define ANSI_COLOR_CYAN    "\x1b[36m"
-#define ANSI_COLOR_RESET   "\x1b[0m"
-#define ANSI_BOLD          "\x1b[1m"
-
-enum console_color_t {
-    CONSOLE_COLOR_DEFAULT=0,
-    CONSOLE_COLOR_PROMPT,
-    CONSOLE_COLOR_USER_INPUT,
-    CONSOLE_COLOR_ERROR
-};
-
-struct console_state {
-    bool multiline_input = false;
-    bool use_color = false;
-    console_color_t color = CONSOLE_COLOR_DEFAULT;
-
-    FILE* out = stdout;
-#if defined (_WIN32)
-    void* hConsole;
-#else
-    FILE* tty = nullptr;
-    termios prev_state;
-#endif
-};
-
-void console_init(console_state & con_st);
-void console_cleanup(console_state & con_st);
-void console_set_color(console_state & con_st, console_color_t color);
-bool console_readline(console_state & con_st, std::string & line);

examples/console.cpp

@@ -0,0 +1,496 @@
+#include "console.h"
+#include <vector>
+#include <iostream>
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <fcntl.h>
+#include <io.h>
+#else
+#include <climits>
+#include <sys/ioctl.h>
+#include <unistd.h>
+#include <wchar.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <signal.h>
+#include <termios.h>
+#endif
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_YELLOW  "\x1b[33m"
+#define ANSI_COLOR_BLUE    "\x1b[34m"
+#define ANSI_COLOR_MAGENTA "\x1b[35m"
+#define ANSI_COLOR_CYAN    "\x1b[36m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+#define ANSI_BOLD          "\x1b[1m"
+
+namespace console {
+
+    //
+    // Console state
+    //
+
+    static bool      advanced_display = false;
+    static bool      simple_io        = true;
+    static display_t current_display  = reset;
+
+    static FILE*     out              = stdout;
+
+#if defined (_WIN32)
+    static void*     hConsole;
+#else
+    static FILE*     tty              = nullptr;
+    static termios   initial_state;
+#endif
+
+    //
+    // Init and cleanup
+    //
+
+    void init(bool use_simple_io, bool use_advanced_display) {
+        advanced_display = use_advanced_display;
+        simple_io = use_simple_io;
+#if defined(_WIN32)
+        // Windows-specific console initialization
+        DWORD dwMode = 0;
+        hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            hConsole = GetStdHandle(STD_ERROR_HANDLE);
+            if (hConsole != INVALID_HANDLE_VALUE && (!GetConsoleMode(hConsole, &dwMode))) {
+                hConsole = nullptr;
+                simple_io = true;
+            }
+        }
+        if (hConsole) {
+            // Enable ANSI colors on Windows 10+
+            if (advanced_display && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
+                SetConsoleMode(hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING);
+            }
+            // Set console output codepage to UTF8
+            SetConsoleOutputCP(CP_UTF8);
+        }
+        HANDLE hConIn = GetStdHandle(STD_INPUT_HANDLE);
+        if (hConIn != INVALID_HANDLE_VALUE && GetConsoleMode(hConIn, &dwMode)) {
+            // Set console input codepage to UTF16
+            _setmode(_fileno(stdin), _O_WTEXT);
+
+            // Set ICANON (ENABLE_LINE_INPUT) and ECHO (ENABLE_ECHO_INPUT)
+            if (simple_io) {
+                dwMode |= ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT;
+            } else {
+                dwMode &= ~(ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT);
+            }
+            if (!SetConsoleMode(hConIn, dwMode)) {
+                simple_io = true;
+            }
+        }
+#else
+        // POSIX-specific console initialization
+        if (!simple_io) {
+            struct termios new_termios;
+            tcgetattr(STDIN_FILENO, &initial_state);
+            new_termios = initial_state;
+            new_termios.c_lflag &= ~(ICANON | ECHO);
+            new_termios.c_cc[VMIN] = 1;
+            new_termios.c_cc[VTIME] = 0;
+            tcsetattr(STDIN_FILENO, TCSANOW, &new_termios);
+
+            tty = fopen("/dev/tty", "w+");
+            if (tty != nullptr) {
+                out = tty;
+            }
+        }
+
+        setlocale(LC_ALL, "");
+#endif
+    }
+
+    void cleanup() {
+        // Reset console display
+        set_display(reset);
+
+#if !defined(_WIN32)
+        // Restore settings on POSIX systems
+        if (!simple_io) {
+            if (tty != nullptr) {
+                out = stdout;
+                fclose(tty);
+                tty = nullptr;
+            }
+            tcsetattr(STDIN_FILENO, TCSANOW, &initial_state);
+        }
+#endif
+    }
+
+    //
+    // Display and IO
+    //
+
+    // Keep track of current display and only emit ANSI code if it changes
+    void set_display(display_t display) {
+        if (advanced_display && current_display != display) {
+            fflush(stdout);
+            switch(display) {
+                case reset:
+                    fprintf(out, ANSI_COLOR_RESET);
+                    break;
+                case prompt:
+                    fprintf(out, ANSI_COLOR_YELLOW);
+                    break;
+                case user_input:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_GREEN);
+                    break;
+                case error:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_RED);
+            }
+            current_display = display;
+            fflush(out);
+        }
+    }
+
+    char32_t getchar32() {
+#if defined(_WIN32)
+        HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
+        wchar_t high_surrogate = 0;
+
+        while (true) {
+            INPUT_RECORD record;
+            DWORD count;
+            if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
+                return WEOF;
+            }
+
+            if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
+                wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
+                if (wc == 0) {
+                    continue;
+                }
+
+                if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+                    high_surrogate = wc;
+                    continue;
+                }
+                if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
+                    if (high_surrogate != 0) { // Check if we have a high surrogate
+                        return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
+                    }
+                }
+
+                high_surrogate = 0; // Reset the high surrogate
+                return static_cast<char32_t>(wc);
+            }
+        }
+#else
+        wchar_t wc = getwchar();
+        if (static_cast<wint_t>(wc) == WEOF) {
+            return WEOF;
+        }
+
+#if WCHAR_MAX == 0xFFFF
+        if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+            wchar_t low_surrogate = getwchar();
+            if ((low_surrogate >= 0xDC00) && (low_surrogate <= 0xDFFF)) { // Check if the next wchar is a low surrogate
+                return (static_cast<char32_t>(wc & 0x03FF) << 10) + (low_surrogate & 0x03FF) + 0x10000;
+            }
+        }
+        if ((wc >= 0xD800) && (wc <= 0xDFFF)) { // Invalid surrogate pair
+            return 0xFFFD; // Return the replacement character U+FFFD
+        }
+#endif
+
+        return static_cast<char32_t>(wc);
+#endif
+    }
+
+    void pop_cursor() {
+#if defined(_WIN32)
+        if (hConsole != NULL) {
+            CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+            GetConsoleScreenBufferInfo(hConsole, &bufferInfo);
+
+            COORD newCursorPosition = bufferInfo.dwCursorPosition;
+            if (newCursorPosition.X == 0) {
+                newCursorPosition.X = bufferInfo.dwSize.X - 1;
+                newCursorPosition.Y -= 1;
+            } else {
+                newCursorPosition.X -= 1;
+            }
+
+            SetConsoleCursorPosition(hConsole, newCursorPosition);
+            return;
+        }
+#endif
+        putc('\b', out);
+    }
+
+    int estimateWidth(char32_t codepoint) {
+#if defined(_WIN32)
+        return 1;
+#else
+        return wcwidth(codepoint);
+#endif
+    }
+
+    int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
+#if defined(_WIN32)
+        CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+        if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) {
+            // go with the default
+            return expectedWidth;
+        }
+        COORD initialPosition = bufferInfo.dwCursorPosition;
+        DWORD nNumberOfChars = length;
+        WriteConsole(hConsole, utf8_codepoint, nNumberOfChars, &nNumberOfChars, NULL);
+
+        CONSOLE_SCREEN_BUFFER_INFO newBufferInfo;
+        GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+
+        // Figure out our real position if we're in the last column
+        if (utf8_codepoint[0] != 0x09 && initialPosition.X == newBufferInfo.dwSize.X - 1) {
+            DWORD nNumberOfChars;
+            WriteConsole(hConsole, &" \b", 2, &nNumberOfChars, NULL);
+            GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+        }
+
+        int width = newBufferInfo.dwCursorPosition.X - initialPosition.X;
+        if (width < 0) {
+            width += newBufferInfo.dwSize.X;
+        }
+        return width;
+#else
+        // We can trust expectedWidth if we've got one
+        if (expectedWidth >= 0 || tty == nullptr) {
+            fwrite(utf8_codepoint, length, 1, out);
+            return expectedWidth;
+        }
+
+        fputs("\033[6n", tty); // Query cursor position
+        int x1;
+        int y1;
+        int x2;
+        int y2;
+        int results = 0;
+        results = fscanf(tty, "\033[%d;%dR", &y1, &x1);
+
+        fwrite(utf8_codepoint, length, 1, tty);
+
+        fputs("\033[6n", tty); // Query cursor position
+        results += fscanf(tty, "\033[%d;%dR", &y2, &x2);
+
+        if (results != 4) {
+            return expectedWidth;
+        }
+
+        int width = x2 - x1;
+        if (width < 0) {
+            // Calculate the width considering text wrapping
+            struct winsize w;
+            ioctl(STDOUT_FILENO, TIOCGWINSZ, &w);
+            width += w.ws_col;
+        }
+        return width;
+#endif
+    }
+
+    void replace_last(char ch) {
+#if defined(_WIN32)
+        pop_cursor();
+        put_codepoint(&ch, 1, 1);
+#else
+        fprintf(out, "\b%c", ch);
+#endif
+    }
+
+    void append_utf8(char32_t ch, std::string & out) {
+        if (ch <= 0x7F) {
+            out.push_back(static_cast<unsigned char>(ch));
+        } else if (ch <= 0x7FF) {
+            out.push_back(static_cast<unsigned char>(0xC0 | ((ch >> 6) & 0x1F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0xFFFF) {
+            out.push_back(static_cast<unsigned char>(0xE0 | ((ch >> 12) & 0x0F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0x10FFFF) {
+            out.push_back(static_cast<unsigned char>(0xF0 | ((ch >> 18) & 0x07)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 12) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else {
+            // Invalid Unicode code point
+        }
+    }
+
+    // Helper function to remove the last UTF-8 character from a string
+    void pop_back_utf8_char(std::string & line) {
+        if (line.empty()) {
+            return;
+        }
+
+        size_t pos = line.length() - 1;
+
+        // Find the start of the last UTF-8 character (checking up to 4 bytes back)
+        for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
+            if ((line[pos] & 0xC0) != 0x80) {
+                break; // Found the start of the character
+            }
+        }
+        line.erase(pos);
+    }
+
+    bool readline_advanced(std::string & line, bool multiline_input) {
+        if (out != stdout) {
+            fflush(stdout);
+        }
+
+        line.clear();
+        std::vector<int> widths;
+        bool is_special_char = false;
+        bool end_of_stream = false;
+
+        char32_t input_char;
+        while (true) {
+            fflush(out); // Ensure all output is displayed before waiting for input
+            input_char = getchar32();
+
+            if (input_char == '\r' || input_char == '\n') {
+                break;
+            }
+
+            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
+                end_of_stream = true;
+                break;
+            }
+
+            if (is_special_char) {
+                set_display(user_input);
+                replace_last(line.back());
+                is_special_char = false;
+            }
+
+            if (input_char == '\033') { // Escape sequence
+                char32_t code = getchar32();
+                if (code == '[' || code == 0x1B) {
+                    // Discard the rest of the escape sequence
+                    while ((code = getchar32()) != (char32_t) WEOF) {
+                        if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
+                            break;
+                        }
+                    }
+                }
+            } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
+                if (!widths.empty()) {
+                    int count;
+                    do {
+                        count = widths.back();
+                        widths.pop_back();
+                        // Move cursor back, print space, and move cursor back again
+                        for (int i = 0; i < count; i++) {
+                            replace_last(' ');
+                            pop_cursor();
+                        }
+                        pop_back_utf8_char(line);
+                    } while (count == 0 && !widths.empty());
+                }
+            } else {
+                int offset = line.length();
+                append_utf8(input_char, line);
+                int width = put_codepoint(line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
+                if (width < 0) {
+                    width = 0;
+                }
+                widths.push_back(width);
+            }
+
+            if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
+                set_display(prompt);
+                replace_last(line.back());
+                is_special_char = true;
+            }
+        }
+
+        bool has_more = multiline_input;
+        if (is_special_char) {
+            replace_last(' ');
+            pop_cursor();
+
+            char last = line.back();
+            line.pop_back();
+            if (last == '\\') {
+                line += '\n';
+                fputc('\n', out);
+                has_more = !has_more;
+            } else {
+                // llama will just eat the single space, it won't act as a space
+                if (line.length() == 1 && line.back() == ' ') {
+                    line.clear();
+                    pop_cursor();
+                }
+                has_more = false;
+            }
+        } else {
+            if (end_of_stream) {
+                has_more = false;
+            } else {
+                line += '\n';
+                fputc('\n', out);
+            }
+        }
+
+        fflush(out);
+        return has_more;
+    }
+
+    bool readline_simple(std::string & line, bool multiline_input) {
+#if defined(_WIN32)
+        std::wstring wline;
+        if (!std::getline(std::wcin, wline)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            GenerateConsoleCtrlEvent(CTRL_C_EVENT, 0);
+            return false;
+        }
+
+        int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), NULL, 0, NULL, NULL);
+        line.resize(size_needed);
+        WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), &line[0], size_needed, NULL, NULL);
+#else
+        if (!std::getline(std::cin, line)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            return false;
+        }
+#endif
+        if (!line.empty()) {
+            char last = line.back();
+            if (last == '/') { // Always return control on '/' symbol
+                line.pop_back();
+                return false;
+            }
+            if (last == '\\') { // '\\' changes the default action
+                line.pop_back();
+                multiline_input = !multiline_input;
+            }
+        }
+        line += '\n';
+
+        // By default, continue input if multiline_input is set
+        return multiline_input;
+    }
+
+    bool readline(std::string & line, bool multiline_input) {
+        set_display(user_input);
+
+        if (simple_io) {
+            return readline_simple(line, multiline_input);
+        }
+        return readline_advanced(line, multiline_input);
+    }
+
+}

examples/console.h

@@ -0,0 +1,19 @@
+// Console functions
+
+#pragma once
+
+#include <string>
+
+namespace console {
+    enum display_t {
+        reset = 0,
+        prompt,
+        user_input,
+        error
+    };
+
+    void init(bool use_simple_io, bool use_advanced_display);
+    void cleanup();
+    void set_display(display_t display);
+    bool readline(std::string & line, bool multiline_input);
+}

examples/embd-input/embd-input-lib.cpp

@@ -30,7 +30,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
+        params.seed = uint32_t(time(NULL));
     }
     fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
 

examples/grammar-parser.cpp

@@ -405,7 +405,7 @@ namespace grammar_parser {
             for (size_t i = 0, end = state.rules.size(); i < end; i++) {
                 // fprintf(file, "%zu: ", i);
                 // print_rule_binary(file, state.rules[i]);
-                print_rule(file, i, state.rules[i], symbol_id_names);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
                 // fprintf(file, "\n");
             }
         } catch (const std::exception & err) {

examples/json-schema-to-grammar.py

@@ -0,0 +1,132 @@
+import argparse
+import json
+import re
+import sys
+
+# whitespace is constrained to a single space char to prevent model "running away" in
+# whitespace. Also maybe improves generation quality?
+SPACE_RULE = '" "?'
+
+PRIMITIVE_RULES = {
+    'boolean': '("true" | "false") space',
+    'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
+    'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space',
+    'string': r''' "\"" (
+        [^"\\] |
+        "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+      )* "\"" space ''',
+    'null': '"null" space',
+}
+
+INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
+GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'}
+
+
+class SchemaConverter:
+    def __init__(self, prop_order):
+        self._prop_order = prop_order
+        self._rules = {'space': SPACE_RULE}
+
+    def _format_literal(self, literal):
+        escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal)
+        )
+        return f'"{escaped}"'
+
+    def _add_rule(self, name, rule):
+        esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
+        if esc_name not in self._rules or self._rules[esc_name] == rule:
+            key = esc_name
+        else:
+            i = 0
+            while f'{esc_name}{i}' in self._rules:
+                i += 1
+            key = f'{esc_name}{i}'
+        self._rules[key] = rule
+        return key
+
+    def visit(self, schema, name):
+        schema_type = schema.get('type')
+        rule_name = name or 'root'
+
+        if 'oneOf' in schema or 'anyOf' in schema:
+            rule = ' | '.join((
+                self.visit(alt_schema, f'{name}{"-" if name else ""}{i}')
+                for i, alt_schema in enumerate(schema.get('oneOf') or schema['anyOf'])
+            ))
+            return self._add_rule(rule_name, rule)
+
+        elif 'const' in schema:
+            return self._add_rule(rule_name, self._format_literal(schema['const']))
+
+        elif 'enum' in schema:
+            rule = ' | '.join((self._format_literal(v) for v in schema['enum']))
+            return self._add_rule(rule_name, rule)
+
+        elif schema_type == 'object' and 'properties' in schema:
+            # TODO: `required` keyword
+            prop_order = self._prop_order
+            prop_pairs = sorted(
+                schema['properties'].items(),
+                # sort by position in prop_order (if specified) then by key
+                key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]),
+            )
+
+            rule = '"{" space'
+            for i, (prop_name, prop_schema) in enumerate(prop_pairs):
+                prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
+                if i > 0:
+                    rule += ' "," space'
+                rule += fr' {self._format_literal(prop_name)} space ":" space {prop_rule_name}'
+            rule += ' "}" space'
+
+            return self._add_rule(rule_name, rule)
+
+        elif schema_type == 'array' and 'items' in schema:
+            # TODO `prefixItems` keyword
+            item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
+            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
+            return self._add_rule(rule_name, rule)
+
+        else:
+            assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
+            return self._add_rule(
+                'root' if rule_name == 'root' else schema_type,
+                PRIMITIVE_RULES[schema_type]
+            )
+
+    def format_grammar(self):
+        return '\n'.join((f'{name} ::= {rule}' for name, rule in self._rules.items()))
+
+
+def main(args_in = None):
+    parser = argparse.ArgumentParser(
+        description='''
+            Generates a grammar (suitable for use in ./main) that produces JSON conforming to a
+            given JSON schema. Only a subset of JSON schema features are supported; more may be
+            added in the future.
+        ''',
+    )
+    parser.add_argument(
+        '--prop-order',
+        default=[],
+        type=lambda s: s.split(','),
+        help='''
+            comma-separated property names defining the order of precedence for object properties;
+            properties not specified here are given lower precedence than those that are, and are
+            sorted alphabetically
+        '''
+    )
+    parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
+    args = parser.parse_args(args_in)
+
+    schema = json.load(sys.stdin if args.schema == '-' else open(args.schema))
+    prop_order = {name: idx for idx, name in enumerate(args.prop_order)}
+    converter = SchemaConverter(prop_order)
+    converter.visit(schema, '')
+    print(converter.format_grammar())
+
+
+if __name__ == '__main__':
+    main()

examples/llama.vim

@@ -0,0 +1,132 @@
+" Requires an already running llama.cpp server
+" To install either copy or symlink to ~/.vim/autoload/llama.vim
+" Then start with either :call llama#doLlamaGen(),
+" or add a keybind to your vimrc such as
+" nnoremap Z :call llama#doLlamaGen()<CR>
+" Similarly, you could add an insert mode keybind with
+" inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
+"
+" g:llama_api_url and g:llama_overrides can be configured in your .vimrc
+" let g:llama_api_url = "192.168.1.10:8080"
+" llama_overrides can also be set through buffer/window scopes. For instance
+" autocmd filetype python let b:llama_overrides = {"temp": 0.2}
+" Could be added to your .vimrc to automatically set a lower temperature when
+" editing a python script
+" Additionally, an override dict can be stored at the top of a file
+" !*{"stop": ["User:"]}
+" Could be added to the start of your chatlog.txt to set the stopping token
+" These parameter dicts are merged together from lowest to highest priority:
+" server default -> g:llama_overrides -> w:llama_overrides ->
+" b:llama_overrides -> in file (!*) overrides
+"
+" Sublists (like logit_bias and stop) are overridden, not merged
+" Example override:
+" !*{"logit_bias": [[13, -5], [2, false]], "temperature": 1, "top_k": 5, "top_p": 0.5, "n_predict": 256, "repeat_last_n": 256, "repeat_penalty": 1.17647}
+if !exists("g:llama_api_url")
+    let g:llama_api_url= "127.0.0.1:8080"
+endif
+if !exists("g:llama_overrides")
+   let g:llama_overrides = {}
+endif
+const s:querydata = {"n_predict": 256, "stop": [ "\n" ], "stream": v:true }
+const s:curlcommand = ['curl','--data-raw', "{\"prompt\":\"### System:\"}", '--silent', '--no-buffer', '--request', 'POST', '--url', g:llama_api_url .. '/completion', '--header', "Content-Type: application/json"]
+let s:linedict = {}
+
+func s:callbackHandler(bufn, channel, msg)
+   if len(a:msg) < 3
+      return
+   elseif a:msg[0] == "d"
+      let l:msg = a:msg[6:-1]
+   else
+      let l:msg = a:msg
+   endif
+   let l:decoded_msg = json_decode(l:msg)
+   let l:newtext = split(l:decoded_msg['content'], "\n", 1)
+   if len(l:newtext) > 0
+      call setbufline(a:bufn, s:linedict[a:bufn], getbufline(a:bufn, s:linedict[a:bufn])[0] .. newtext[0])
+   else
+      echo "nothing genned"
+   endif
+   if len(newtext) > 1
+      let l:failed = appendbufline(a:bufn, s:linedict[a:bufn], newtext[1:-1])
+      let s:linedict[a:bufn] = s:linedict[a:bufn] + len(newtext)-1
+   endif
+   if has_key(l:decoded_msg, "stop") && l:decoded_msg.stop
+       echo "Finished generation"
+   endif
+endfunction
+
+func llama#doLlamaGen()
+   if exists("b:job")
+      if job_status(b:job) == "run"
+         call job_stop(b:job)
+         return
+      endif
+   endif
+
+   let l:cbuffer = bufnr("%")
+   let s:linedict[l:cbuffer] = line('$')
+   let l:buflines = getbufline(l:cbuffer, 1, 1000)
+   let l:querydata = copy(s:querydata)
+   call extend(l:querydata, g:llama_overrides)
+   if exists("w:llama_overrides")
+      call extend(l:querydata, w:llama_overrides)
+   endif
+   if exists("b:llama_overrides")
+      call extend(l:querydata, b:llama_overrides)
+   endif
+   if l:buflines[0][0:1] == '!*'
+      let l:userdata = json_decode(l:buflines[0][2:-1])
+      call extend(l:querydata, l:userdata)
+      let l:buflines = l:buflines[1:-1]
+   endif
+   let l:querydata.prompt = join(l:buflines, "\n")
+   let l:curlcommand = copy(s:curlcommand)
+   let l:curlcommand[2] = json_encode(l:querydata)
+   let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
+endfunction
+
+" Echos the tokkenization of the provided string , or cursor to end of word
+" Onus is placed on the user to include the preceding space
+func llama#tokenizeWord(...)
+    if (a:0 > 0)
+        let l:input = a:1
+    else
+        exe "normal \"*ye"
+        let l:input = @*
+    endif
+    let l:querydata = {"content": l:input}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": function("s:tokenizeWordCallback", [l:input])})
+endfunction
+
+func s:tokenizeWordCallback(plaintext, channel, msg)
+    echo '"' .. a:plaintext ..'" - ' .. string(json_decode(a:msg).tokens)
+endfunction
+
+
+" Echos the token count of the entire buffer (or provided string)
+" Example usage :echo llama#tokenCount()
+func llama#tokenCount(...)
+    if (a:0 > 0)
+        let l:buflines = a:1
+    else
+        let l:buflines = getline(1,1000)
+        if l:buflines[0][0:1] == '!*'
+            let l:buflines = l:buflines[1:-1]
+        endif
+        let l:buflines = join(l:buflines, "\n")
+    endif
+    let l:querydata = {"content": l:buflines}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": "s:tokenCountCallback"})
+endfunction
+
+func s:tokenCountCallback(channel, msg)
+    let resp = json_decode(a:msg)
+    echo len(resp.tokens)
+endfunction

examples/llm.vim

@@ -1,3 +1,5 @@
+" Basic plugin example
+
 function! Llm()
 
   let url = "http://127.0.0.1:8080/completion"
@@ -16,8 +18,10 @@ function! Llm()
   " Extract the content field from the response
   let content = json_decode(response).content
 
+  let split_newlines = split(content, '\n', 1)
+
   " Insert the content at the cursor position
-  call setline(line('.'), getline('.') . content)
+  call setline(line('.'), [ getline('.') . split_newlines[0] ] + split_newlines[1:])
 endfunction
 
 command! Llm call Llm()

examples/main/README.md

@@ -140,6 +140,12 @@ The `--ctx-size` option allows you to set the size of the prompt context used by
 
 -   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
 
+### Extended Context Size
+
+Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
+
+- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+
 ### Keep Prompt
 
 The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained.
@@ -202,9 +208,9 @@ Example usage: `--top-p 0.95`
 
 -   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
 
-Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. The method adjusts the logits (token probabilities) by raising them to the power of the parameter z. A higher value of z (e.g., 2.0) will further suppress less likely tokens from the tail of the distribution, while a value of 1.0 disables the effect of TFS. By setting the parameter z, you can control how much the probabilities of less likely tokens are reduced.
+Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens, and thus disables the effect of TFS.
 
-Example usage: `--tfs 2.0`
+Example usage: `--tfs 0.95`
 
 ### Locally Typical Sampling
 

examples/main/main.cpp

@@ -4,6 +4,7 @@
 #endif
 
 #include "common.h"
+#include "console.h"
 #include "llama.h"
 #include "build-info.h"
 #include "grammar-parser.h"
@@ -35,9 +36,11 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static console_state con_st;
-static llama_context ** g_ctx;
+#if defined(GGML_USE_KOMPUTE)
+#include "ggml-vulkan.h"
+#endif
 
+static llama_context ** g_ctx;
 static bool is_interacting = false;
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
@@ -46,7 +49,7 @@ void sigint_handler(int signo) {
         if (!is_interacting) {
             is_interacting=true;
         } else {
-            console_cleanup(con_st);
+            console::cleanup();
             printf("\n");
             llama_print_timings(*g_ctx);
             _exit(130);
@@ -64,10 +67,8 @@ int main(int argc, char ** argv) {
 
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
-    con_st.use_color = params.use_color;
-    con_st.multiline_input = params.multiline_input;
-    console_init(con_st);
-    atexit([]() { console_cleanup(con_st); });
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
 
     if (params.perplexity) {
         printf("\n************\n");
@@ -121,6 +122,10 @@ int main(int argc, char ** argv) {
     llama_context * ctx_guidance = NULL;
     g_ctx = &ctx;
 
+#if defined(GGML_USE_KOMPUTE)
+    ggml_vk_init_device(0, "gpu");
+#endif
+
     // load the model and apply lora adapter, if any
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (params.cfg_scale > 1.f) {
@@ -373,7 +378,7 @@ int main(int argc, char ** argv) {
 
     if (params.interactive) {
         const char *control_message;
-        if (con_st.multiline_input) {
+        if (params.multiline_input) {
             control_message = " - To return control to LLaMa, end your input with '\\'.\n"
                               " - To return control without starting a new line, end your input with '/'.\n";
         } else {
@@ -401,7 +406,7 @@ int main(int argc, char ** argv) {
     int n_past_guidance    = 0;
 
     // the first thing we will do is to output the prompt, so set color accordingly
-    console_set_color(con_st, CONSOLE_COLOR_PROMPT);
+    console::set_display(console::prompt);
 
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
@@ -422,9 +427,9 @@ int main(int argc, char ** argv) {
             // Ensure the input doesn't exceed the context size by truncating embd if necessary.
             if ((int)embd.size() > max_embd_size) {
                 auto skipped_tokens = embd.size() - max_embd_size;
-                console_set_color(con_st, CONSOLE_COLOR_ERROR);
+                console::set_display(console::error);
                 printf("<<input too long: skipped %zu token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
-                console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+                console::set_display(console::reset);
                 fflush(stdout);
                 embd.resize(max_embd_size);
             }
@@ -667,7 +672,7 @@ int main(int argc, char ** argv) {
         }
         // reset color to default if we there is no pending user input
         if (input_echo && (int)embd_inp.size() == n_consumed) {
-            console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+            console::set_display(console::reset);
         }
 
         // if not currently processing queued inputs;
@@ -693,7 +698,7 @@ int main(int argc, char ** argv) {
                     if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) {
                         if (params.interactive) {
                             is_interacting = true;
-                            console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
+                            console::set_display(console::user_input);
                         }
                         is_antiprompt = true;
                         fflush(stdout);
@@ -714,7 +719,7 @@ int main(int argc, char ** argv) {
 
                     is_interacting = true;
                     printf("\n");
-                    console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
+                    console::set_display(console::user_input);
                     fflush(stdout);
                 } else if (params.instruct) {
                     is_interacting = true;
@@ -739,12 +744,12 @@ int main(int argc, char ** argv) {
                 std::string line;
                 bool another_line = true;
                 do {
-                    another_line = console_readline(con_st, line);
+                    another_line = console::readline(line, params.multiline_input);
                     buffer += line;
                 } while (another_line);
 
                 // done taking input, reset color
-                console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+                console::set_display(console::reset);
 
                 // Add tokens to embd only if the input buffer is non-empty
                 // Entering a empty line lets the user pass control back

examples/perplexity/perplexity.cpp

@@ -121,8 +121,23 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     printf("\n");
 }
 
-void perplexity_lines(llama_context * ctx, const gpt_params & params) {
-    // Calculates perplexity over each line of the prompt
+void hellaswag_score(llama_context * ctx, const gpt_params & params) {
+    // Calculates hellaswag score (acc_norm) from prompt
+    //
+    // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
+    // All used data fields are preprocessed as in https://github.com/EleutherAI/lm-evaluation-harness/blob/df3da98c5405deafd519c2ddca52bb7c3fe36bef/lm_eval/tasks/hellaswag.py#L62-L68
+    //
+    // All 10042 tasks should be extracted to keep the results standardized like other implementations.
+    //
+    // Datafile layout:
+    // ['??'] denotes json fields
+    // 6 lines per task:
+    // ['activity_label'] + ": " +['ctx']  - The first part of the query, the context
+    // ['label'] - The index the best common sense ending aka gold ending
+    // ['endings'][0] - Endings added to the first part of the query
+    // ['endings'][1]
+    // ['endings'][2]
+    // ['endings'][3]
 
     std::vector<std::string> prompt_lines;
     std::istringstream strstream(params.prompt);
@@ -132,63 +147,149 @@ void perplexity_lines(llama_context * ctx, const gpt_params & params) {
         prompt_lines.push_back(line);
     }
 
+    if( prompt_lines.size() % 6 != 0) {
+        fprintf(stderr, "%s : number of lines in prompt not a multiple of 6.\n", __func__);
+        return;
+    }
+
+    size_t hs_task_count = prompt_lines.size()/6;
+    fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
+
+    // This is needed as usual for LLaMA models
+    bool prepend_bos = true;
+
+    // Number of tasks to use when computing the score
+    if ( params.hellaswag_tasks < hs_task_count  ) {
+        hs_task_count = params.hellaswag_tasks;
+    }
+
+    // The tasks should be randomized so the score stabilizes quickly.
+    bool randomize_tasks = true;
+
+    // The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now
+    std::mt19937 rng(1);
+
+    // Dataholder for hellaswag tasks
+    struct hs_data_t {
+        std::string context;
+        size_t gold_ending_idx;
+        std::string ending[4];
+        size_t ending_logprob_count[4];
+        double ending_logprob[4];
+    };
+
+    fprintf(stderr, "%s : selecting %zu %s tasks.\n", __func__, hs_task_count, (randomize_tasks?"randomized":"the first")  );
+
+    // Select and read data from prompt lines
+    hs_data_t *hs_data = new hs_data_t[hs_task_count];
+    for (size_t i=0; i < hs_task_count; i++) {
+        size_t idx = i;
+
+        // Select a random example of those left in the prompt
+        if (randomize_tasks) {
+            std::uniform_int_distribution<size_t> dist(0, prompt_lines.size()/6-1 ) ;
+            idx = dist(rng);
+        }
+
+        hs_data[i].context = prompt_lines[idx*6];
+        hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
+        for (size_t j=0; j < 4; j++) {
+            hs_data[i].ending[j] = " " + prompt_lines[idx*6+2+j];
+        }
+
+        // Delete the selected random example from the prompt
+        if (randomize_tasks) {
+            prompt_lines.erase( std::next(prompt_lines.begin(),idx*6)  , std::next(prompt_lines.begin(),idx*6+6) );
+        }
+    }
+
+    fprintf(stderr, "%s : calculating hellaswag score over selected tasks.\n", __func__);
+    printf("\ntask\tacc_norm\n");
+
+    double acc = 0.0f;
     const int n_vocab = llama_n_vocab(ctx);
 
-    int counttotal   = 0;
-    size_t n_lines = prompt_lines.size();
+    for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
 
-    double nll = 0.0;
+        // Tokenize the context to count tokens
+        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos);
+        size_t context_size = context_embd.size();
 
-    fprintf(stderr, "%s: calculating perplexity over %lu lines\n", __func__, n_lines);
+        for (size_t ending_idx=0;ending_idx<4;ending_idx++) {
 
-    printf("\nLine\tPPL line\tPPL cumulative\n");
+            // Tokenize the query
+            std::vector<int> query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[ending_idx], prepend_bos);
+            size_t query_size = query_embd.size();
 
-    for (size_t i = 0; i < n_lines; ++i) {
+            // Stop if query wont fit the ctx window
+            if (query_size > (size_t)params.n_ctx) {
+                fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
+                return;
+            }
 
-        // Tokenize and insert BOS at start
-        std::vector<int> batch_embd = ::llama_tokenize(ctx, prompt_lines[i], true);
+            // Speedup small evaluations by evaluating atleast 32 tokens
+            if (query_size < 32) {
+                query_embd.resize(32);
+            }
 
-        size_t batch_size  = batch_embd.size();
+            // Evaluate the query
+            if (llama_eval(ctx, query_embd.data(), query_embd.size(), 0, params.n_threads)) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return;
+            }
 
-        // Stop if line is too long
-        if( batch_size > (size_t)params.n_ctx ) {
-            fprintf(stderr, "%s : tokens in line %lu > n_ctxl\n", __func__, i);
-            return;
+            const auto query_logits = llama_get_logits(ctx);
+            std::vector<float> logits;
+            logits.insert(logits.end(), query_logits, query_logits + query_size * n_vocab);
+
+            hs_data[task_idx].ending_logprob_count[ending_idx] = 0;
+            hs_data[task_idx].ending_logprob[ending_idx] = 0.0f;
+
+            // Calculate the logprobs over the ending
+            for (size_t j = context_size-1; j < query_size - 1; j++) {
+                // Calculate probability of next token, given the previous ones.
+                const std::vector<float> tok_logits(
+                    logits.begin() + (j + 0) * n_vocab,
+                    logits.begin() + (j + 1) * n_vocab);
+
+                const float prob = softmax(tok_logits)[query_embd[ j + 1]];
+
+                hs_data[task_idx].ending_logprob[ending_idx] += std::log(prob);
+                hs_data[task_idx].ending_logprob_count[ending_idx]++;
+            }
+
+            // Calculate the mean token logprob for acc_norm
+            hs_data[task_idx].ending_logprob[ending_idx] /= hs_data[task_idx].ending_logprob_count[ending_idx];
+
+
+//            printf("task %lu, ending %lu, whole_len %lu, context_len %lu, ending_logprob_count %lu, ending_logprob %.4f\n",
+//                task_idx,ending_idx,whole_size,context_size, hs_data[task_idx].ending_logprob_count[ending_idx], hs_data[task_idx].ending_logprob[ending_idx] );
         }
 
-        if (llama_eval(ctx, batch_embd.data(), batch_size, 0, params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return;
+        // Find the ending with maximum logprob
+        size_t ending_logprob_max_idx = -1;
+        double ending_logprob_max_val = -INFINITY;
+        for (size_t j=0; j < 4; j++) {
+            if (hs_data[task_idx].ending_logprob[j] > ending_logprob_max_val) {
+                ending_logprob_max_idx = j;
+                ending_logprob_max_val =  hs_data[task_idx].ending_logprob[j];
+            }
         }
 
-        const auto batch_logits = llama_get_logits(ctx);
-        std::vector<float> logits;
-        logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+//        printf("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_data[task_idx].gold_ending_idx);
 
-        double nllline = 0.0;
-        int countline = 0;
-
-        // Perplexity over second half of the line
-        for (size_t j = batch_size/2; j < batch_size - 1; ++j) {
-            // Calculate probability of next token, given the previous ones.
-            const std::vector<float> tok_logits(
-                logits.begin() + (j + 0) * n_vocab,
-                logits.begin() + (j + 1) * n_vocab);
-
-            const float prob = softmax(tok_logits)[batch_embd[ j + 1]];
-
-            nllline += -std::log(prob);
-            ++countline;
+        // If the gold ending got the maximum logprobe add one accuracy point
+        if (ending_logprob_max_idx == hs_data[task_idx].gold_ending_idx) {
+            acc += 1.0;
         }
 
-        nll += nllline;
-        counttotal += countline;
-
-        // perplexity is e^(average negative log-likelihood)
-        printf("%lu\t%.8lf\t%.8lf\n", i + 1, std::exp(nllline/countline), std::exp(nll / counttotal) );
+        // Print the accumulated accuracy mean x 100
+        printf("%zu\t%.8lf\n",task_idx+1, acc/double(task_idx+1)*100.0);
         fflush(stdout);
     }
 
+    delete [] hs_data;
+
     printf("\n");
 }
 
@@ -240,8 +341,8 @@ int main(int argc, char ** argv) {
                 params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
     }
 
-    if (params.perplexity_lines) {
-        perplexity_lines(ctx, params);
+    if (params.hellaswag) {
+        hellaswag_score(ctx, params);
     } else {
         perplexity(ctx, params);
     }

examples/save-load-state/save-load-state.cpp

@@ -26,6 +26,7 @@ int main(int argc, char ** argv) {
     auto lparams = llama_context_default_params();
 
     lparams.n_ctx     = params.n_ctx;
+    lparams.n_gqa     = params.n_gqa;
     lparams.seed      = params.seed;
     lparams.f16_kv    = params.memory_f16;
     lparams.use_mmap  = params.use_mmap;

examples/server-llama2-13B.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+
+set -e
+
+cd "$(dirname "$0")/.." || exit
+
+# Specify the model you want to use here:
+MODEL="${MODEL:-./models/llama-2-13b-chat.ggmlv3.q5_K_M.bin}"
+PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat-system.txt}
+
+# Adjust to the number of CPU cores you want to use.
+N_THREAD="${N_THREAD:-12}"
+
+# Note: you can also override the generation options by specifying them on the command line:
+GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 4096 --batch-size 1024}"
+
+
+# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
+./server $GEN_OPTIONS \
+  --model "$MODEL" \
+  --threads "$N_THREAD" \
+  --rope-freq-scale 1.0 \
+  "$@"
+
+# I used this to test the model with mps, but omitted it from the general purpose. If you want to use it, just specify it on the command line.
+# -ngl 1 \

examples/server/README.md

@@ -163,7 +163,7 @@ node .
 
     `content`: Set the text to tokenize.
 
-    Note that the special `BOS` token is not added in fron of the text and also a space character is not inserted automatically as it is for `/completion`.
+    Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
 
 -   **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
 

examples/server/chat-llama2.sh

@@ -0,0 +1,109 @@
+#!/bin/bash
+
+API_URL="${API_URL:-http://127.0.0.1:8080}"
+
+CHAT=(
+    "Hello, Assistant."
+    "Hello. How may I help you today?"
+)
+
+INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
+
+trim() {
+    shopt -s extglob
+    set -- "${1##+([[:space:]])}"
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+trim_trailing() {
+    shopt -s extglob
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+format_prompt() {
+    if [[ "${#CHAT[@]}" -eq 0 ]]; then
+        echo -n "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>"
+    else
+        LAST_INDEX=$(( ${#CHAT[@]} - 1 ))
+        echo -n "${CHAT[$LAST_INDEX]}\n[INST] $1 [/INST]"
+    fi
+}
+
+tokenize() {
+    curl \
+        --silent \
+        --request POST \
+        --url "${API_URL}/tokenize" \
+        --header "Content-Type: application/json" \
+        --data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
+    | jq '.tokens[]'
+}
+
+N_KEEP=$(tokenize "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>" | wc -l)
+
+chat_completion() {
+    PROMPT="$(trim_trailing "$(format_prompt "$1")")"
+    DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
+        prompt: .,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: $n_keep,
+        n_predict: 1024,
+        stop: ["[INST]"],
+        stream: true
+    }')"
+
+    # Create a temporary file to hold the Python output
+    TEMPFILE=$(mktemp)
+
+    exec 3< <(curl \
+        --silent \
+        --no-buffer \
+        --request POST \
+        --url "${API_URL}/completion" \
+        --header "Content-Type: application/json" \
+        --data-raw "${DATA}")
+
+    python -c "
+import json
+import sys
+
+answer = ''
+while True:
+    line = sys.stdin.readline()
+    if not line:
+        break
+    if line.startswith('data: '):
+        json_content = line[6:].strip()
+        content = json.loads(json_content)['content']
+        sys.stdout.write(content)
+        sys.stdout.flush()
+        answer += content
+
+answer = answer.rstrip('\n')
+
+# Write the answer to the temporary file
+with open('$TEMPFILE', 'w') as f:
+    f.write(answer)
+    " <&3
+
+    exec 3<&-
+
+    # Read the answer from the temporary file
+    ANSWER=$(cat $TEMPFILE)
+
+    # Clean up the temporary file
+    rm $TEMPFILE
+
+    printf "\n"
+
+    CHAT+=("$1" "$(trim "$ANSWER")")
+}
+
+while true; do
+    echo -en "\033[0;32m"  # Green color
+    read -r -e -p "> " QUESTION
+    echo -en "\033[0m"  # Reset color
+    chat_completion "${QUESTION}"
+done

examples/server/completion.js.hpp

@@ -87,289 +87,342 @@ unsigned char completion_js[] = {
   0x20, 0x54, 0x65, 0x78, 0x74, 0x44, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72,
   0x28, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63,
   0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d,
-  0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20,
-  0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72,
-  0x2e, 0x72, 0x65, 0x61, 0x64, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c,
-  0x74, 0x2e, 0x64, 0x6f, 0x6e, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x73, 0x65, 0x20, 0x61,
-  0x6e, 0x73, 0x77, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68,
-  0x65, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6d, 0x75, 0x6c, 0x74, 0x69,
-  0x70, 0x6c, 0x65, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x20, 0x6f, 0x66,
-  0x3a, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5c, 0x6e, 0x20, 0x77, 0x69,
-  0x74, 0x68, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x61, 0x6c, 0x77, 0x61,
-  0x79, 0x73, 0x20, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x74, 0x20, 0x61,
-  0x73, 0x20, 0x61, 0x20, 0x6b, 0x65, 0x79, 0x2e, 0x20, 0x69, 0x6e, 0x20,
-  0x6f, 0x75, 0x72, 0x20, 0x63, 0x61, 0x73, 0x65, 0x20, 0x77, 0x65, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6d, 0x61, 0x69,
-  0x6e, 0x6c, 0x79, 0x20, 0x63, 0x61, 0x72, 0x65, 0x20, 0x61, 0x62, 0x6f,
-  0x75, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3a,
-  0x20, 0x6b, 0x65, 0x79, 0x20, 0x68, 0x65, 0x72, 0x65, 0x2c, 0x20, 0x77,
-  0x68, 0x69, 0x63, 0x68, 0x20, 0x77, 0x65, 0x20, 0x65, 0x78, 0x70, 0x65,
-  0x63, 0x74, 0x20, 0x61, 0x73, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74,
-  0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65,
-  0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x28, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x61,
-  0x72, 0x73, 0x65, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20,
-  0x65, 0x76, 0x65, 0x6e, 0x74, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61,
-  0x64, 0x64, 0x20, 0x74, 0x68, 0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20,
-  0x3d, 0x20, 0x2f, 0x5e, 0x28, 0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73,
-  0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x20, 0x6f, 0x66, 0x20,
-  0x74, 0x65, 0x78, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x41, 0x6c,
-  0x6c, 0x28, 0x72, 0x65, 0x67, 0x65, 0x78, 0x29, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x5b, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d,
-  0x20, 0x3d, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6e, 0x63, 0x65, 0x20,
-  0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20, 0x74, 0x68, 0x69, 0x73,
-  0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70,
-  0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x6a, 0x75, 0x73,
-  0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x74, 0x68, 0x65,
-  0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x64, 0x61, 0x74,
-  0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53,
-  0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x20, 0x2b, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x79,
-  0x69, 0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79,
-  0x69, 0x65, 0x6c, 0x64, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69,
-  0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20, 0x61, 0x20, 0x73,
-  0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x20, 0x66, 0x72,
-  0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2c, 0x20, 0x77,
-  0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b,
-  0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
-  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e,
-  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
-  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
-  0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72,
-  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e,
-  0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72,
-  0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20, 0x63,
-  0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61, 0x6d,
-  0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72, 0x74,
-  0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e,
-  0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x65,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f,
-  0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74,
-  0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b,
-  0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20, 0x74,
-  0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x79,
-  0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x63, 0x72,
-  0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f,
-  0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f,
-  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72,
-  0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65,
-  0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d, 0x20, 0x66,
-  0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
-  0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f, 0x0a,
-  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28,
-  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76,
-  0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28,
-  0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20, 0x28,
-  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75,
-  0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63,
-  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x2e,
-  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70,
-  0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c,
-  0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
-  0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
-  0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b,
-  0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20,
-  0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
-  0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45,
-  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28,
-  0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c,
-  0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d,
-  0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72,
-  0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c,
-  0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c,
-  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e,
-  0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e,
-  0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
-  0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74,
-  0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74,
-  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61,
-  0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77,
-  0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74,
-  0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20,
-  0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x29, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72,
-  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e,
-  0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76,
-  0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74,
-  0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74,
-  0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64,
+  0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
+  0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f,
+  0x20, 0x42, 0x75, 0x66, 0x66, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20,
+  0x70, 0x61, 0x72, 0x74, 0x69, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x72, 0x65,
+  0x61, 0x64, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20,
+  0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65,
+  0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75,
+  0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c,
+  0x65, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69,
+  0x74, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61,
+  0x64, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f,
+  0x6e, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x20, 0x41, 0x64, 0x64, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x6c,
+  0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x61, 0x74, 0x61,
+  0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x75, 0x72, 0x72,
+  0x65, 0x6e, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66,
+  0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d,
+  0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x2b, 0x20,
+  0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f,
+  0x64, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x20, 0x69, 0x66,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x63, 0x68,
+  0x61, 0x72, 0x61, 0x63, 0x74, 0x65, 0x72, 0x20, 0x69, 0x73, 0x20, 0x61,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65,
+  0x42, 0x72, 0x65, 0x61, 0x6b, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74,
+  0x2e, 0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x5c,
+  0x6e, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x20, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
+  0x65, 0x74, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x5c,
+  0x6e, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x20, 0x49, 0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65,
+  0x78, 0x74, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x6e, 0x27, 0x74, 0x20, 0x65,
+  0x6e, 0x64, 0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x61, 0x20, 0x6c, 0x69,
+  0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x2c, 0x20, 0x74, 0x68,
+  0x65, 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20,
+  0x6c, 0x69, 0x6e, 0x65, 0x20, 0x69, 0x73, 0x20, 0x69, 0x6e, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x53, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74,
+  0x20, 0x69, 0x6e, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72,
+  0x20, 0x74, 0x6f, 0x20, 0x62, 0x65, 0x20, 0x61, 0x64, 0x64, 0x65, 0x64,
+  0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6e, 0x65, 0x78, 0x74,
+  0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61,
+  0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x21, 0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69,
+  0x6e, 0x65, 0x42, 0x72, 0x65, 0x61, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
+  0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x2e,
+  0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20,
+  0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76,
+  0x65, 0x72, 0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x68, 0x61, 0x76,
+  0x65, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65,
+  0x61, 0x6b, 0x20, 0x61, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x6e,
+  0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x50, 0x61, 0x72, 0x73,
+  0x65, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76,
+  0x65, 0x6e, 0x74, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64,
+  0x20, 0x74, 0x68, 0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20,
+  0x2f, 0x5e, 0x28, 0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e,
+  0x2a, 0x29, 0x24, 0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x69, 0x6e,
+  0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63,
+  0x68, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x2e, 0x65, 0x78,
+  0x65, 0x63, 0x28, 0x6c, 0x69, 0x6e, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b,
+  0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20,
+  0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69,
+  0x6e, 0x63, 0x65, 0x20, 0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20,
+  0x74, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x2e, 0x63, 0x70, 0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73,
+  0x20, 0x6a, 0x75, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e,
+  0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75,
+  0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d,
+  0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d,
+  0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c,
+  0x64, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20,
+  0x61, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e,
+  0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72,
+  0x2c, 0x20, 0x77, 0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72,
+  0x65, 0x61, 0x6b, 0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61,
+  0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
+  0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
+  0x73, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
   0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69,
-  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20,
-  0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63,
-  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69,
-  0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
-  0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63,
-  0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43,
-  0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22,
-  0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20,
-  0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e,
-  0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76,
-  0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74,
-  0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64, 0x6f, 0x6e,
-  0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
-  0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
-  0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
-  0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b,
-  0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20,
-  0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65,
-  0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6d,
-  0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e,
-  0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20, 0x6e,
-  0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x73,
-  0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f, 0x2f, 0x0a,
-  0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c,
-  0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70,
-  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28,
-  0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e,
-  0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
-  0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a,
-  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f,
-  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d,
-  0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x6d,
-  0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64,
-  0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74,
-  0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d,
-  0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
+  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20,
+  0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62,
+  0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20,
+  0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61,
+  0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72,
+  0x74, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
+  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20,
+  0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
+  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20,
+  0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20,
+  0x79, 0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x63,
+  0x72, 0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f,
+  0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f,
+  0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76,
+  0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d, 0x20,
+  0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c,
+  0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72,
+  0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20,
+  0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63,
+  0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
+  0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
+  0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
   0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20,
   0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d,
   0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x50,
-  0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63,
-  0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c, 0x20, 0x72,
-  0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74,
-  0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
-  0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
-  0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72,
-  0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
-  0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
-  0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
-  0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20,
-  0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x65,
-  0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x2f,
-  0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70, 0x72, 0x65,
-  0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f, 0x0a, 0x65,
-  0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74,
-  0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
-  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62,
-  0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f,
-  0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61,
-  0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47, 0x65, 0x74,
-  0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, 0x69,
-  0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x68, 0x65,
-  0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54, 0x68, 0x69,
-  0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75, 0x6c, 0x20,
-  0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x20,
-  0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x20,
-  0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73,
-  0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20,
-  0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20,
+  0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20,
+  0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
+  0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f,
+  0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20,
+  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
+  0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f,
+  0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61,
+  0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70,
+  0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65,
+  0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e,
+  0x74, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c,
+  0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x29,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73,
+  0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67, 0x65,
+  0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74,
+  0x74, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65,
+  0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e,
+  0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74,
+  0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73,
+  0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74,
+  0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74,
+  0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20,
+  0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x22, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b,
+  0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73,
+  0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64, 0x6f,
+  0x6e, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69,
+  0x6c, 0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x20, 0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x29,
+  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
+  0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
+  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65,
+  0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
+  0x65, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78, 0x74,
+  0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20,
+  0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20,
+  0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a,
+  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28,
+  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e,
+  0x28, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72,
+  0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72, 0x0a,
+  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
+  0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f,
+  0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
+  0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a,
+  0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f,
+  0x6d, 0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d,
+  0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
+  0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20,
+  0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77, 0x20,
+  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79, 0x6e,
+  0x63, 0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c, 0x20,
+  0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74,
+  0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e,
+  0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70,
+  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c,
+  0x76, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68,
+  0x20, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x28,
+  0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a,
+  0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70, 0x72,
+  0x65, 0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f, 0x0a,
+  0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
+  0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c, 0x6c,
+  0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20,
+  0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62,
+  0x61, 0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47, 0x65,
+  0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20,
+  0x69, 0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x68,
+  0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54, 0x68,
+  0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75, 0x6c,
+  0x20, 0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74,
+  0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64, 0x20,
+  0x73, 0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72,
+  0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d,
+  0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67, 0x65,
+  0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74,
+  0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x61,
+  0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x22,
+  0x2f, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x22,
+  0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20,
+  0x72, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
   0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
-  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x61, 0x77,
-  0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x22, 0x2f,
-  0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x22, 0x29,
-  0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20, 0x72,
-  0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67,
-  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
+  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
 };
-unsigned int completion_js_len = 4462;
+unsigned int completion_js_len = 5099;

examples/server/public/completion.js

@@ -43,6 +43,7 @@ export async function* llama(prompt, params = {}, config = {}) {
   const decoder = new TextDecoder();
 
   let content = "";
+  let leftover = ""; // Buffer for partially read lines
 
   try {
     let cont = true;
@@ -53,29 +54,47 @@ export async function* llama(prompt, params = {}, config = {}) {
         break;
       }
 
-      // sse answers in the form multiple lines of: value\n with data always present as a key. in our case we
-      // mainly care about the data: key here, which we expect as json
-      const text = decoder.decode(result.value);
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value);
 
-      // parse all sse events and add them to result
-      const regex = /^(\S+):\s(.*)$/gm;
-      for (const match of text.matchAll(regex)) {
-        result[match[1]] = match[2]
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n');
+
+      // Split the text into lines
+      let lines = text.split('\n');
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop();
+      } else {
+        leftover = ""; // Reset leftover if we have a line break at the end
       }
 
-      // since we know this is llama.cpp, let's just decode the json in data
-      result.data = JSON.parse(result.data);
-      content += result.data.content;
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const line of lines) {
+        const match = regex.exec(line);
+        if (match) {
+          result[match[1]] = match[2]
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data);
+            content += result.data.content;
 
-      // yield
-      yield result;
+            // yield
+            yield result;
 
-      // if we got a stop token from server, we will break here
-      if (result.data.stop) {
-        if (result.data.generation_settings) {
-          generation_settings = result.data.generation_settings;
+            // if we got a stop token from server, we will break here
+            if (result.data.stop) {
+              if (result.data.generation_settings) {
+                generation_settings = result.data.generation_settings;
+              }
+              cont = false;
+              break;
+            }
+          }
         }
-        break;
       }
     }
   } catch (e) {

examples/server/public/index.html

@@ -3,12 +3,11 @@
 <head>
   <meta charset="UTF-8">
   <meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1" />
+  <meta name="color-scheme" content="light dark">
   <title>llama.cpp - chat</title>
 
   <style>
     body {
-      background-color: #fff;
-      color: #000;
       font-family: system-ui;
       font-size: 90%;
     }
@@ -283,8 +282,9 @@
 
       useEffect(() => {
         // scroll to bottom (if needed)
-        if (container.current && container.current.scrollHeight <= container.current.scrollTop + container.current.offsetHeight + 300) {
-          container.current.scrollTo(0, container.current.scrollHeight)
+        const parent = container.current.parentElement;
+        if (parent && parent.scrollHeight <= parent.scrollTop + parent.offsetHeight + 300) {
+          parent.scrollTo(0, parent.scrollHeight)
         }
       }, [messages])
 

examples/server/server.cpp

@@ -631,6 +631,9 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
     fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
+    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
+    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
+    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
 #endif
     fprintf(stdout, "  -m FNAME, --model FNAME\n");
     fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
@@ -827,7 +830,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 }
             }
 #else
-            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.", {});
+            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--low-vram" || arg == "-lv")
@@ -835,7 +838,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
 #ifdef GGML_USE_CUBLAS
             params.low_vram = true;
 #else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
+#endif // GGML_USE_CUBLAS
+        }
+        else if (arg == "--mul-mat-q" || arg == "-mmq")
+        {
+#ifdef GGML_USE_CUBLAS
+            params.mul_mat_q = true;
+#else
+            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--main-gpu" || arg == "-mg")
@@ -1263,7 +1274,11 @@ int main(int argc, char **argv)
                 sink.done();
                 return true;
             };
-            res.set_chunked_content_provider("text/event-stream", chunked_content_provider);
+            const auto on_complete = [&](bool) {
+                llama.mutex.unlock();
+            };
+            lock.release();
+            res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
         } });
 
     svr.Get("/model.json", [&llama](const Request &, Response &res)

examples/simple/simple.cpp

@@ -123,7 +123,7 @@ int main(int argc, char ** argv)
         // Evaluate the tokens :
         //---------------------------------
 
-        if ( llama_eval( ctx , tokens_list.data() , tokens_list.size() , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
+        if ( llama_eval( ctx , tokens_list.data() , int(tokens_list.size()) , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
         {
             fprintf( stderr,  "%s : failed to eval\n" , __func__ );
             return 1;

ggml-alloc.c

@@ -0,0 +1,541 @@
+#include "ggml-alloc.h"
+#include "ggml.h"
+#include <assert.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED(x) (void)(x)
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF printf
+#define AT_PRINTF(...) ((void)0)
+
+struct hash_node {
+    struct ggml_tensor * t;
+    int n_children;
+    int n_views;
+};
+
+static size_t hash(void * p) {
+    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
+}
+
+static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
+    size_t h = hash(t);
+
+    // linear probing
+    size_t i = h;
+    while (hash_table[i].t != NULL) {
+        if (hash_table[i].t == t) {
+            return &hash_table[i];
+        }
+        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
+        if (i == h) {
+            // hash table is full
+            GGML_ASSERT(false);
+        }
+    }
+
+    hash_table[i].t = t;
+    return &hash_table[i];
+}
+
+// TODO: GGML_PAD ?
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+struct free_block {
+    void * addr;
+    size_t size;
+};
+
+#define MAX_FREE_BLOCKS 128
+
+struct ggml_allocr {
+    void * data;
+    size_t size;
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+    size_t max_size;
+    bool measure;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct ggml_tensor * allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == NULL) {
+            alloc->allocated_tensors[i] = tensor;
+            return;
+        }
+    }
+    GGML_ASSERT(!"out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == tensor ||
+            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
+            alloc->allocated_tensors[i] = NULL;
+            return;
+        }
+    }
+    printf("tried to free tensor %s not found\n", tensor->name);
+    GGML_ASSERT(!"tensor not found");
+}
+#endif
+
+
+static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    return ggml_nbytes(tensor);
+
+    UNUSED(alloc);
+}
+
+void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    AT_PRINTF("block %d\n", best_fit_block);
+
+    if (best_fit_block == -1) {
+        fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
+                __func__, size, max_avail);
+        GGML_ASSERT(!"not enough space in the buffer");
+        return;
+    }
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    void * addr = block->addr;
+    block->addr = (char*)block->addr + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    tensor->data = addr;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, tensor);
+    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    if (cur_max > alloc->max_size) {
+        printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i]) {
+                printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
+            }
+        }
+        printf("\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
+}
+
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+
+    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+        // the tensor was not allocated in this buffer
+        // this can happen because the graph allocator will try to free weights and other tensors from different buffers
+        // the easiest way to deal with this is just to ignore it
+        return;
+    }
+
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if ((char*)block->addr + block->size == ptr) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if ((char*)ptr + size == block->addr) {
+            block->addr = ptr;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].addr = ptr;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+}
+
+void ggml_allocr_reset(struct ggml_allocr * alloc) {
+    alloc->n_free_blocks = 1;
+    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
+    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
+    alloc->free_blocks[0].size = alloc->size - align_offset;
+}
+
+struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ data,
+        /*.size          = */ size,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ false,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+// address and size of the buffer when measuring
+// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
+static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
+static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+
+struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ MEASURE_BASE_ADDR,
+        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ true,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+void ggml_allocr_free(struct ggml_allocr * alloc) {
+    free(alloc);
+}
+
+bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
+    return alloc->measure;
+}
+
+//////////// compute graph allocator
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
+           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
+    switch (t->op) {
+        case GGML_OP_PERMUTE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+            return t->src[0];
+        case GGML_OP_CPY:
+            return t->src[1];
+        default:
+            return NULL;
+    }
+}
+
+static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
+    struct ggml_tensor * parent = t;
+    do {
+        parent = get_view_parent(parent);
+    } while (ggml_is_view(parent));
+    return parent;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_ACC:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SET:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_CONT:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
+    struct hash_node * ht = alloc->hash_table;
+    if (node->data == NULL) {
+        if (ggml_is_view(node)) {
+            size_t offset;
+            switch(node->op) {
+                case GGML_OP_VIEW:
+                    memcpy(&offset, node->op_params, sizeof(size_t));
+                    node->data = (char *) node->src[0]->data + offset;
+                    break;
+                case GGML_OP_PERMUTE:
+                case GGML_OP_RESHAPE:
+                case GGML_OP_TRANSPOSE:
+                    node->data = node->src[0]->data;
+                    break;
+                case GGML_OP_CPY:
+                    node->data = node->src[1]->data;
+                    break;
+                default:
+                    GGML_ASSERT(!"unknown view op");
+                    break;
+            }
+        } else {
+            // see if we can reuse a parent's buffer (inplace)
+            if (ggml_op_can_inplace(node->op)) {
+                for (int i = 0; i < GGML_MAX_SRC; i++) {
+                    struct ggml_tensor * parent = node->src[i];
+                    if (parent == NULL) {
+                        break;
+                    }
+                    struct hash_node * p_hn = hash_get(ht, parent);
+                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
+                        if (ggml_is_view(parent)) {
+                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct hash_node * view_src_hn = hash_get(ht, view_src);
+                            if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                                // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
+                                // the parent's data that it will need later (same layout requirement). the problem is that then
+                                // we cannot free the tensor because the original address of the allocation is lost.
+                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
+                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
+                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                                node->data = parent->data;
+                                return;
+                            }
+                        }
+                        else {
+                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                            node->data = parent->data;
+                        }
+                        return;
+                    }
+                }
+            }
+            ggml_allocr_alloc(alloc, node);
+        }
+    }
+}
+
+static size_t ggml_allocator_alloc_graph_tensors_n(
+    struct ggml_allocr * alloc,
+    struct ggml_cgraph ** graphs, int n_graphs,
+    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
+
+    // reset hash table
+    struct hash_node * ht = alloc->hash_table;
+    memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
+
+    // count number of children and views
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            if (ggml_is_view(node)) {
+                struct ggml_tensor * view_src = get_view_source(node);
+                hash_get(ht, view_src)->n_views += 1;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                hash_get(ht, parent)->n_children += 1;
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
+        // graph inputs are allocated first to ensure that they are not overwritten by each other
+        if (inputs != NULL && inputs[g] != NULL) {
+            for (int i = 0; inputs[g][i] != NULL; i++) {
+                struct ggml_tensor * input = inputs[g][i];
+                AT_PRINTF("input: %s\n", input->name);
+                allocate_node(alloc, input);
+            }
+        }
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            // allocate parents (leafs)
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                allocate_node(alloc, parent);
+            }
+
+            // allocate node
+            allocate_node(alloc, node);
+
+            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                AT_PRINTF("%s", parent->name);
+                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                    AT_PRINTF(", ");
+                }
+            }
+            AT_PRINTF("\n");
+
+            // update parents
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                struct hash_node * p_hn = hash_get(ht, parent);
+                p_hn->n_children -= 1;
+
+                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
+
+                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = get_view_source(parent);
+                        struct hash_node * view_src_hn = hash_get(ht, view_src);
+                        view_src_hn->n_views -= 1;
+                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
+                        if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, view_src);
+                        }
+                    }
+                    else {
+                        if (parent->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, parent);
+                        }
+                    }
+                }
+            }
+            AT_PRINTF("\n");
+        }
+        // free graph outputs here that wouldn't be freed otherwise because they have no children
+        if (outputs != NULL && outputs[g] != NULL) {
+            for (int i = 0; outputs[g][i] != NULL; i++) {
+                struct ggml_tensor * output = outputs[g][i];
+                AT_PRINTF("output: %s\n", output->name);
+                ggml_allocator_free_tensor(alloc, output);
+            }
+        }
+    }
+
+    return alloc->max_size;
+}
+
+size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
+    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+}

ggml-alloc.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+
+GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
+GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
+
+GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
+GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
+GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+
+
+#ifdef  __cplusplus
+}
+#endif

ggml-cuda.h

@@ -27,6 +27,7 @@ void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
 void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
 void   ggml_cuda_set_scratch_size(size_t scratch_size);
 void   ggml_cuda_free_scratch(void);
 bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);

ggml-metal.m

@@ -7,6 +7,11 @@
 #import <Metal/Metal.h>
 #import <MetalPerformanceShaders/MetalPerformanceShaders.h>
 
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
 #ifdef GGML_METAL_NDEBUG
 #define metal_printf(...)
 #else
@@ -15,6 +20,8 @@
 
 #define UNUSED(x) (void)(x)
 
+#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+
 struct ggml_metal_buffer {
     const char * name;
 
@@ -36,7 +43,7 @@ struct ggml_metal_context {
     int n_buffers;
     struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
 
-    int concur_list[GGML_MAX_NODES];
+    int concur_list[GGML_MAX_CONCUR];
     int concur_list_len;
 
     // custom kernels
@@ -370,15 +377,15 @@ void ggml_metal_graph_find_concurrency(
         struct ggml_metal_context * ctx,
         struct ggml_cgraph * gf) {
     int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
-    int nodes_unused[GGML_MAX_NODES];
+    int nodes_unused[GGML_MAX_CONCUR];
 
-    for (int i = 0; i < GGML_MAX_NODES; i++) {ctx->concur_list[i] = 0;}
-    for (int i = 0; i < gf->n_nodes; i++) {nodes_unused[i] = 1;}
+    for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
+    for (int i = 0; i < gf->n_nodes;     i++) { nodes_unused[i]     = 1; }
     ctx->concur_list_len = 0;
 
-    int n_left = gf->n_nodes;
-    int n_start = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
-    int level_pos = 0;  // at ctx->concur_list, the last layer (level) ends at level_pos
+    int n_left    = gf->n_nodes;
+    int n_start   = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
+    int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
 
     while (n_left > 0) {
         // number of nodes at a layer (that can be issued concurrently)
@@ -386,28 +393,40 @@ void ggml_metal_graph_find_concurrency(
         for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
             if (nodes_unused[i]) {
                 // if the requirements for gf->nodes[i] are satisfied
-                int exe_flag=1;
+                int exe_flag = 1;
+
                 // scan all srcs
                 for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
                     struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
                     if (src_cur) {
                         // if is leaf nodes it's satisfied.
-                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {continue;}
+                        // TODO: ggml_is_leaf()
+                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
+                            continue;
+                        }
 
                         // otherwise this src should be the output from previous nodes.
                         int is_found = 0;
+
                         // scan 2*search_depth back because we inserted barrier.
-                        for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
-                            if (gf->nodes[ctx->concur_list[j]] == src_cur) {is_found = 1; break;}
+                        //for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
+                        for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
+                            if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
+                                is_found = 1;
+                                break;
+                            }
+                        }
+                        if (is_found == 0) {
+                            exe_flag = 0;
+                            break;
                         }
-                        if (is_found == 0) {exe_flag = 0; break;}
                     }
                 }
                 if (exe_flag) {
                     // check if nodes[i]'s data will be overwritten by a node before nodes[i].
                     // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
                     int64_t data_start = (int64_t) gf->nodes[i]->data;
-                    int64_t length = (int64_t) ggml_nbytes(gf->nodes[i]);
+                    int64_t length     = (int64_t) ggml_nbytes(gf->nodes[i]);
                     for (int j = n_start; j < i; j++) {
                         if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
                                             && gf->nodes[j]->op != GGML_OP_VIEW \
@@ -416,9 +435,9 @@ void ggml_metal_graph_find_concurrency(
                             if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
                                 ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
                                 continue;
-                            } else {
-                                exe_flag = 0;
                             }
+
+                            exe_flag = 0;
                         }
                     }
                 }
@@ -435,11 +454,13 @@ void ggml_metal_graph_find_concurrency(
         ctx->concur_list[level_pos + concurrency] = -1;
         ctx->concur_list_len++;
         // jump all sorted nodes at nodes_bak
-        while (!nodes_unused[n_start]) {n_start++;}
+        while (!nodes_unused[n_start]) {
+            n_start++;
+        }
         level_pos += concurrency + 1;
     }
 
-    if (ctx->concur_list_len > GGML_MAX_NODES) {
+    if (ctx->concur_list_len > GGML_MAX_CONCUR) {
         fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
     }
 }
@@ -453,7 +474,7 @@ void ggml_metal_graph_compute(
     // else fallback to serial dispatch
     MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
 
-    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_NODES;
+    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
 
     const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
     edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
@@ -718,7 +739,8 @@ void ggml_metal_graph_compute(
                             // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
 
                             GGML_ASSERT(ne00 == ne10);
-                            GGML_ASSERT(ne02 == ne12);
+                            // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
+                            GGML_ASSERT(ne03 == ne13);
 
                             if (ggml_is_contiguous(src0) &&
                                 ggml_is_contiguous(src1) &&
@@ -746,11 +768,11 @@ void ggml_metal_graph_compute(
                                     initWithDevice:ctx->device transposeLeft:false transposeRight:true
                                         resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
 
-                                // we need to do ne02 multiplications
+                                // we need to do ne12 multiplications
                                 // TODO: is there a way to do this in parallel - currently very slow ..
                                 // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
-                                for (int64_t i02 = 0; i02 < ne02; ++i02) {
-                                    size_t offs_src0_cur = offs_src0 + i02*nb02;
+                                for (int64_t i02 = 0; i02 < ne12; ++i02) {
+                                    size_t offs_src0_cur = offs_src0 + i02/(ne12/ne02)*nb02; // gqa not used for now
                                     size_t offs_src1_cur = offs_src1 + i02*nb12;
                                     size_t offs_dst_cur  = offs_dst  + i02*nb2;
 
@@ -772,8 +794,6 @@ void ggml_metal_graph_compute(
                                 switch (src0t) {
                                     case GGML_TYPE_F16:
                                         {
-                                            GGML_ASSERT(ne02 == ne12);
-
                                             nth0 = 64;
                                             nth1 = 1;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
@@ -853,16 +873,18 @@ void ggml_metal_graph_compute(
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
                                 [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
                                 [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
-                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
-                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
-                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
-                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
-                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
-                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
-                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
-                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
-                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+                                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
+                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                     src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {

ggml-metal.metal

@@ -509,11 +509,13 @@ kernel void kernel_mul_mat_f16_f32(
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
+        constant   int64_t & ne02,
         constant  uint64_t & nb00,
         constant  uint64_t & nb01,
         constant  uint64_t & nb02,
         constant   int64_t & ne10,
         constant   int64_t & ne11,
+        constant   int64_t & ne12,
         constant  uint64_t & nb10,
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
@@ -529,7 +531,7 @@ kernel void kernel_mul_mat_f16_f32(
     const int64_t r1 = tgpig.y;
     const int64_t im = tgpig.z;
 
-    device const half  * x = (device const half  *) (src0 + r0*nb01 + im*nb02);
+    device const half  * x = (device const half  *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
     device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
 
     sum[tpitg.x] = 0.0f;
@@ -552,6 +554,7 @@ kernel void kernel_mul_mat_f16_f32(
     }
 }
 
+
 kernel void kernel_alibi_f32(
         device const float * src0,
         device       float * dst,

ggml-vulkan.h

@@ -0,0 +1,63 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#pragma once
+
+#include <cstddef>
+#include <vector>
+#include <string>
+
+struct ggml_kompute_context;
+
+namespace vk {
+    class DeviceMemory;
+    class Buffer;
+};
+
+struct ggml_vk_memory {
+    void *data = nullptr;
+    size_t size = 0;
+    vk::DeviceMemory *primaryMemory = nullptr;
+    vk::Buffer *primaryBuffer = nullptr;
+    vk::DeviceMemory *stagingMemory = nullptr;
+    vk::Buffer *stagingBuffer = nullptr;
+};
+
+struct ggml_vk_device {
+    int index = 0;
+    int type = 0;           // same as VkPhysicalDeviceType
+    size_t heapSize = 0;
+    std::string name;
+    std::string vendor;
+};
+
+std::vector<ggml_vk_device> ggml_vk_available_devices(size_t memoryRequired);
+bool ggml_vk_init_device(size_t memoryRequired, const std::string &device);
+bool ggml_vk_init_device(const ggml_vk_device &device);
+bool ggml_vk_init_device(int device);
+bool ggml_vk_free_device();
+bool ggml_vk_has_vulkan();
+bool ggml_vk_has_device();
+ggml_vk_device ggml_vk_current_device();
+struct ggml_kompute_context * ggml_vk_init(void);
+bool ggml_vk_has_h2d_all(struct ggml_kompute_context * ctx);
+void ggml_vk_free(struct ggml_kompute_context * ctx);
+size_t ggml_vk_aligned_offset(size_t offset);
+ggml_vk_memory ggml_vk_allocate(size_t size);
+void ggml_vk_free_memory(ggml_vk_memory &memory);
+
+void ggml_vk_add_buffer(
+    struct ggml_kompute_context * ctx,
+    const char * name,
+    const ggml_vk_memory &memory);
+
+void ggml_vk_h2d_all(struct ggml_kompute_context * ctx);
+void ggml_vk_d2h_all(struct ggml_kompute_context * ctx);
+void ggml_vk_h2d_tensor(struct ggml_kompute_context * ctx, struct ggml_tensor * t);
+void ggml_vk_d2h_tensor(struct ggml_kompute_context * ctx, struct ggml_tensor * t);
+void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf);

ggml.h

@@ -183,6 +183,15 @@
 #    define GGML_API
 #endif
 
+// TODO: support for clang
+#ifdef __GNUC__
+#    define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define GGML_DEPRECATED(func, hint) func
+#endif
+
 #include <stdint.h>
 #include <stddef.h>
 #include <stdbool.h>
@@ -374,6 +383,10 @@ extern "C" {
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_MAP_CUSTOM1_F32,
+        GGML_OP_MAP_CUSTOM2_F32,
+        GGML_OP_MAP_CUSTOM3_F32,
+
         GGML_OP_MAP_CUSTOM1,
         GGML_OP_MAP_CUSTOM2,
         GGML_OP_MAP_CUSTOM3,
@@ -570,6 +583,8 @@ extern "C" {
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
 
+    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
 
@@ -1170,7 +1185,18 @@ extern "C" {
             int                   mode,
             int                   n_ctx);
 
-    // custom RoPE, in-place, returns view(a)
+    // custom RoPE
+    GGML_API struct ggml_tensor * ggml_rope_custom(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx,
+            float                 freq_base,
+            float                 freq_scale);
+
+    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1229,7 +1255,7 @@ extern "C" {
 
     // conv_1d with padding = half
     // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
-    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
+    GGML_API struct ggml_tensor * ggml_conv_1d_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
@@ -1242,7 +1268,7 @@ extern "C" {
         GGML_OP_POOL_COUNT,
     };
 
-    GGML_API struct ggml_tensor* ggml_pool_1d(
+    GGML_API struct ggml_tensor * ggml_pool_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             enum ggml_op_pool     op,
@@ -1250,7 +1276,7 @@ extern "C" {
             int                   s0, // stride
             int                   p0); // padding
 
-    GGML_API struct ggml_tensor* ggml_pool_2d(
+    GGML_API struct ggml_tensor * ggml_pool_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             enum ggml_op_pool     op,
@@ -1304,15 +1330,6 @@ extern "C" {
             int                   h0,
             int                   w);
 
-    // custom operators
-
-    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-
     GGML_API struct ggml_tensor * ggml_unary(
             struct ggml_context * ctx,
              struct ggml_tensor * a,
@@ -1323,63 +1340,138 @@ extern "C" {
         struct ggml_tensor  * a,
         enum ggml_unary_op op);
 
-    GGML_API struct ggml_tensor * ggml_map_unary_f32(
+    // custom operators
+
+    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
+    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+
+    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3_inplace instead");
+
+    // custom operators v2
+
+    typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
+
+    #define GGML_N_TASKS_MAX -1
+
+    GGML_API struct ggml_tensor * ggml_map_custom1(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
 
     // loss function
 

grammars/json.gbnf

@@ -1,29 +1,25 @@
-# Grammar for subset of JSON - doesn't support full string or number syntax
-
-root  ::= object
-value ::= object | array | string | number | boolean | "null"
+root   ::= object
+value  ::= object | array | string | number | ("true" | "false" | "null") ws
 
 object ::=
   "{" ws (
             string ":" ws value
     ("," ws string ":" ws value)*
-  )? "}"
+  )? "}" ws
 
 array  ::=
   "[" ws (
             value
     ("," ws value)*
-  )? "]"
+  )? "]" ws
 
-string  ::=
+string ::=
   "\"" (
     [^"\\] |
     "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
   )* "\"" ws
 
-# Only plain integers currently
-number  ::= "-"? [0-9]+ ws
-boolean ::= ("true" | "false") ws
+number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
 
 # Optional space: by convention, applied in this grammar after literal chars when allowed
 ws ::= ([ \t\n] ws)?

k_quants.c

@@ -39,6 +39,8 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -1353,7 +1355,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -1421,7 +1423,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
 
         // sumf += -dmin * summs in 32bits*8
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(_mm256_set_m128i(summs_1, summs_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
 
         const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
         const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
@@ -1493,7 +1495,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // sumf += dall * isum - dmin * summs in 32bits
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -1644,8 +1646,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * smin;
 
         const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
-        const __m256i q2_0 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 2), q2bits), m3);
-        const __m256i q2_1 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
+        const __m256i q2_0 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits), m3);
+        const __m256i q2_1 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
 
         const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
         const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
@@ -1709,10 +1711,10 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
         const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
 
-        const __m256i p_0 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
-        const __m256i p_1 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
-        const __m256i p_2 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
-        const __m256i p_3 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
+        const __m256i p_0 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
+        const __m256i p_1 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
+        const __m256i p_2 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
+        const __m256i p_3 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
 
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
@@ -1917,7 +1919,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         // high bit
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
@@ -2128,7 +2130,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // multiply with block scale and accumulate
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2303,13 +2305,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         aux16[0] = a & 0x0f0f;
         aux16[1] = (a >> 4) & 0x0f0f;
 
-        const __m256i scale_0 = _mm256_set_m128i(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
-        const __m256i scale_1 = _mm256_set_m128i(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
+        const __m256i scale_0 = MM256_SET_M128I(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
+        const __m256i scale_1 = MM256_SET_M128I(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
 
         memcpy(&aux64, x[i].hmask, 8);
 
         const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
-        __m256i q3h_0 = _mm256_set_m128i(_mm_srli_epi16(haux, 2), haux);
+        __m256i q3h_0 = MM256_SET_M128I(_mm_srli_epi16(haux, 2), haux);
         __m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4);
         q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2);
         q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2);
@@ -2318,7 +2320,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
 
         // prepare low and high bits
-        const __m256i q3aux  = _mm256_set_m128i(_mm_srli_epi16(q3bits, 2), q3bits);
+        const __m256i q3aux  = MM256_SET_M128I(_mm_srli_epi16(q3bits, 2), q3bits);
         const __m256i q3l_0 = _mm256_and_si256(q3aux, m3);
         const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3);
 
@@ -2429,7 +2431,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
         p16_0 = _mm_add_epi32(p16_0, p16_2);
         p16_1 = _mm_add_epi32(p16_1, p16_3);
-        __m256i p16 = _mm256_set_m128i(p16_1, p16_0);
+        __m256i p16 = MM256_SET_M128I(p16_1, p16_0);
 
         // multiply with block scale and accumulate
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
@@ -2620,7 +2622,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -2727,7 +2729,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2968,11 +2970,11 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
         const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
         const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_1, p32_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_1, p32_0))), acc);
 
         const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
         const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_3, p32_2))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_3, p32_2))), acc);
 
     }
 
@@ -3160,7 +3162,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * _mm_extract_epi32(hsum, 0);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
         __m256i hmask = mone;
@@ -3299,7 +3301,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -3462,13 +3464,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
         const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
 
-        const __m256i scale_l = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
-        const __m256i scale_h = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
+        const __m256i scale_l = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
+        const __m256i scale_h = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
 
         int64_t aux64;
         memcpy(&aux64, x[i].qh, 8);
         const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64);
-        const __m256i haux256 = _mm256_set_m128i(_mm_srli_epi16(haux128, 2), haux128);
+        const __m256i haux256 = MM256_SET_M128I(_mm_srli_epi16(haux128, 2), haux128);
 
         const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4);
         const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4);
@@ -3543,7 +3545,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
         const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
 
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_set_m128i(dot_1, dot_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(dot_1, dot_0))), acc);
 
     }
 
@@ -3925,7 +3927,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
         }
 
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -4083,8 +4085,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
         const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
 
-        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
-        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
+        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
+        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
 
         const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
         const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1);
@@ -4177,7 +4179,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
         sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
         sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
 
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(_mm256_set_m128i(sumi_1, sumi_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi_1, sumi_0))), acc);
     }
 
     *s = hsum_float_8(acc);

kompute/.ccls

@@ -0,0 +1,27 @@
+
+%clang
+
+-fdeclspec
+-fms-extensions
+-Wall
+-Wextra
+-std=c++17
+
+%h -x
+%h c++-header
+
+-DDEBUG=1
+-DKOMPUTE_INCLUDE_FOR_SYNTAX
+
+-I/usr/include/python3.6/
+-I./python/pybind11/include/
+
+-I./build/_deps/vulkan_header-src/include/
+-I./build/_deps/spdlog-src/include/
+-I./build/_deps/googletest-src/googletest/include/
+-I./build/_deps/fmt-src/include/
+
+-I./src/include/
+-I./build/src/shaders/glsl/
+-I./build/test/shaders/glsl/
+-I./test/utils/

kompute/.clang-format

@@ -0,0 +1,5 @@
+---
+BasedOnStyle: Mozilla
+IndentWidth: 4
+
+...

kompute/.dockerignore

@@ -0,0 +1,4 @@
+build/*
+examples/*
+docker-builders/
+swiftshader/

kompute/.github/workflows/cpp_examples.yml

@@ -0,0 +1,58 @@
+name: C++ Tests
+
+on:
+  push:
+    branches: [ master ]
+  pull_request:
+    branches: [ master ]
+
+jobs:
+  array-multiplication-example:
+    runs-on: ubuntu-latest
+    container: axsauze/kompute-builder:0.4
+    env:
+      VK_ICD_FILENAMES: "/swiftshader/vk_swiftshader_icd.json"
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+      with:
+        submodules: false
+    - name: "[Release g++] Build & Test"
+      uses: KomputeProject/action-cmake-build@master
+      with:
+        build-dir: ${{github.workspace}}/examples/array_multiplication/build
+        source-dir: ${{github.workspace}}/examples/array_multiplication
+        cc: gcc
+        cxx: g++
+        build-type: Debug
+        run-test: false
+        ctest-options: -V
+        configure-options: -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=ON KOMPUTE_OPT_FROM_SOURCE=ON
+        build-options: --parallel # Given we don't build too many resources we can leverage parallel
+    - name: Run tests
+      run: ./examples/array_multiplication/build/src/kompute_array_mult
+
+  logistc-regression-example:
+    runs-on: ubuntu-latest
+    container: axsauze/kompute-builder:0.4
+    env:
+      VK_ICD_FILENAMES: "/swiftshader/vk_swiftshader_icd.json"
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+      with:
+        submodules: false
+    - name: "[Release g++] Build & Test"
+      uses: KomputeProject/action-cmake-build@master
+      with:
+        build-dir: ${{github.workspace}}/examples/logistic_regression/build
+        source-dir: ${{github.workspace}}/examples/logistic_regression
+        cc: gcc
+        cxx: g++
+        build-type: Debug
+        run-test: false
+        ctest-options: -V
+        configure-options: -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=ON KOMPUTE_OPT_FROM_SOURCE=ON
+        build-options: --parallel # Given we don't build too many resources we can leverage parallel
+    - name: Run tests
+      run: ./examples/logistic_regression/build/src/kompute_logistic_regression

kompute/.github/workflows/cpp_tests.yml

@@ -0,0 +1,104 @@
+name: C++ Tests
+
+on:
+  push:
+    branches: [ master ]
+  pull_request:
+    branches: [ master ]
+
+jobs:
+  cpp-tests-debug-with-debug-layers:
+    runs-on: ubuntu-latest
+    container: axsauze/kompute-builder:0.4
+    env:
+      VK_ICD_FILENAMES: "/swiftshader/vk_swiftshader_icd.json"
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+      with:
+        submodules: false
+    - name: "[Release g++] Build & Test"
+      uses: KomputeProject/action-cmake-build@master
+      with:
+        build-dir: ${{github.workspace}}/build
+        source-dir: ${{github.workspace}}
+        cc: gcc
+        cxx: g++
+        build-type: Debug
+        run-test: false
+        ctest-options: -V
+        configure-options: -DKOMPUTE_OPT_BUILD_TESTS=ON -DKOMPUTE_OPT_DISABLE_VK_DEBUG_LAYERS=OFF -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=ON
+    - name: Run tests
+      run: make mk_run_tests
+
+  cpp-tests-release-with-debug-layers:
+    runs-on: ubuntu-latest
+    container: axsauze/kompute-builder:0.4
+    env:
+      VK_ICD_FILENAMES: "/swiftshader/vk_swiftshader_icd.json"
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+      with:
+        submodules: false
+    - name: "[Release g++] Build & Test"
+      uses: KomputeProject/action-cmake-build@master
+      with:
+        build-dir: ${{github.workspace}}/build
+        source-dir: ${{github.workspace}}
+        cc: gcc
+        cxx: g++
+        build-type: Release
+        run-test: false
+        ctest-options: -V
+        configure-options: -DKOMPUTE_OPT_BUILD_TESTS=ON -DKOMPUTE_OPT_DISABLE_VK_DEBUG_LAYERS=OFF -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=ON
+    - name: Run tests
+      run: make mk_run_tests
+
+  cpp-tests-debug-without-debug-layers:
+    runs-on: ubuntu-latest
+    container: axsauze/kompute-builder:0.4
+    env:
+      VK_ICD_FILENAMES: "/swiftshader/vk_swiftshader_icd.json"
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+      with:
+        submodules: false
+    - name: "[Release g++] Build & Test"
+      uses: KomputeProject/action-cmake-build@master
+      with:
+        build-dir: ${{github.workspace}}/build
+        source-dir: ${{github.workspace}}
+        cc: gcc
+        cxx: g++
+        build-type: Debug
+        run-test: false
+        ctest-options: -V
+        configure-options: -DKOMPUTE_OPT_BUILD_TESTS=ON -DKOMPUTE_OPT_DISABLE_VK_DEBUG_LAYERS=ON -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=ON
+    - name: Run tests
+      run: make mk_run_tests
+  
+  cpp-tests-release-without-debug-layers:
+    runs-on: ubuntu-latest
+    container: axsauze/kompute-builder:0.4
+    env:
+      VK_ICD_FILENAMES: "/swiftshader/vk_swiftshader_icd.json"
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+      with:
+        submodules: false
+    - name: "[Release g++] Build & Test"
+      uses: KomputeProject/action-cmake-build@master
+      with:
+        build-dir: ${{github.workspace}}/build
+        source-dir: ${{github.workspace}}
+        cc: gcc
+        cxx: g++
+        build-type: Release
+        run-test: false
+        ctest-options: -V
+        configure-options: -DKOMPUTE_OPT_BUILD_TESTS=ON -DKOMPUTE_OPT_DISABLE_VK_DEBUG_LAYERS=ON -DKOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER=ON
+    - name: Run tests
+      run: make mk_run_tests

kompute/.github/workflows/python_tests.yml

@@ -0,0 +1,28 @@
+name: Python Tests
+
+on:
+  push:
+    branches: [ master ]
+  pull_request:
+    branches: [ master ]
+
+jobs:
+  python-tests:
+    runs-on: ubuntu-latest
+    container: axsauze/kompute-builder:0.4
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v3
+      with:
+        submodules: false
+    - name: Install Python Requirements
+      run: pip3 install --user -r python/test/requirements-dev.txt
+    - name: Python Build
+      env:
+        KOMPUTE_PYTHON_NUM_PARALLEL_THREADS: 2
+        KOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER: ON
+      run: pip3 install --user . -v
+    - name: Python run Tests
+      run: |
+        export VK_ICD_FILENAMES=/swiftshader/vk_swiftshader_icd.json
+        make test_python

kompute/CMakeLists.txt

@@ -0,0 +1,189 @@
+# SPDX-License-Identifier: Apache-2.0
+
+cmake_minimum_required(VERSION 3.20)
+project(kompute VERSION 0.8.1 LANGUAGES CXX)
+
+set(CMAKE_CXX_STANDARD 14)
+
+# Only change the folder behavior if kompute is not a subproject
+if(${CMAKE_PROJECT_NAME} STREQUAL ${PROJECT_NAME})
+    set_property(GLOBAL PROPERTY USE_FOLDERS ON)
+    set_property(GLOBAL PROPERTY PREDEFINED_TARGETS_FOLDER "CMake")
+    set(EXECUTABLE_OUTPUT_PATH ${CMAKE_BINARY_DIR}/bin)
+    set(LIBRARY_OUTPUT_PATH ${CMAKE_BINARY_DIR}/lib)
+endif()
+
+# Avoid the dll boilerplate code for windows
+set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+set(CMAKE_CXX_STANDARD 14)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+set(CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake;${CMAKE_MODULE_PATH}")
+
+set(KOMPUTE_LIBRARIES kompute CACHE INTERNAL "")
+
+# ####################################################
+# Options
+# ####################################################
+macro(kompute_option OPTION_NAME OPTION_TEXT OPTION_DEFAULT)
+    option(${OPTION_NAME} ${OPTION_TEXT} ${OPTION_DEFAULT})
+
+    if(DEFINED ENV{${OPTION_NAME}})
+        # Allow overriding the option through an environment variable
+        set(${OPTION_NAME} $ENV{${OPTION_NAME}})
+    endif()
+
+    if(${OPTION_NAME})
+        add_definitions(-D${OPTION_NAME})
+    endif()
+
+    message(STATUS "  ${OPTION_NAME}: ${${OPTION_NAME}}")
+endmacro()
+
+macro(kompute_log_level OPTION_NAME OPTION_TEXT OPTION_DEFAULT)
+    set(${OPTION_NAME} ${OPTION_DEFAULT} CACHE STRING ${OPTION_TEXT})
+    set_property(CACHE ${OPTION_NAME} PROPERTY STRINGS "Trace" "Debug" "Info" "Warn" "Error" "Critical" "Default" "Off")
+
+    if(DEFINED ENV{${OPTION_NAME}})
+        # Allow setting the option through an environment variable
+        set(${OPTION_NAME} $ENV{${OPTION_NAME}})
+    endif()
+
+    if(${OPTION_NAME})
+        add_definitions(-D${OPTION_NAME})
+    endif()
+
+    # Allow disabling logging completely and prevent linking against it:
+    if(${KOMPUTE_OPT_LOG_LEVEL} STREQUAL "Off")
+        set(${OPTION_NAME}_DISABLED ON)
+        add_compile_definitions(${OPTION_NAME}_DISABLED=1)
+    endif()
+
+    message(STATUS "  ${OPTION_NAME}: ${${OPTION_NAME}}")
+endmacro()
+
+macro(kompute_option_string OPTION_NAME OPTION_TEXT OPTION_DEFAULT)
+    set(${OPTION_NAME} ${OPTION_DEFAULT} CACHE STRING ${OPTION_TEXT})
+
+    if(DEFINED ENV{${OPTION_NAME}})
+        # Allow setting the option through an environment variable
+        set(${OPTION_NAME} $ENV{${OPTION_NAME}})
+    endif()
+
+    if(${OPTION_NAME})
+        add_definitions(-D${OPTION_NAME})
+    endif()
+
+    message(STATUS "  ${OPTION_NAME}: ${${OPTION_NAME}}")
+endmacro()
+
+message(STATUS "General purpose GPU compute framework built on Vulkan")
+message(STATUS "=======================================================")
+
+# Build options
+kompute_log_level(KOMPUTE_OPT_LOG_LEVEL "Internally we use Spdlog or fmt for logging, depending on the value of 'KOMPUTE_OPT_USE_SPDLOG'. The log level used can be changed here. Possible values: 'Trace', 'Debug', 'Info', 'Warn', 'Error', 'Critical', 'Off', 'Default'. If set to 'Off' logging will be deactivated completely. If set to 'Default', the log level will be set to 'Info' for release builds and 'Debug' else." "Off")
+kompute_option(KOMPUTE_OPT_USE_SPDLOG "If enabled, logging via KP_LOG_<DEBUG, INFO, etc...> will happen through Spdlog instead of plan fmt." OFF)
+kompute_option(KOMPUTE_OPT_DISABLE_VK_DEBUG_LAYERS "Explicitly disable debug layers even on debug." ON)
+kompute_option(KOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK "Whether to check if your driver supports the Vulkan Header version you are linking against. This might be useful in case you build shared on a different system than you run later." OFF)
+kompute_option(KOMPUTE_OPT_BUILD_SHADERS "Rebuilds all compute shaders during compilation and does not use the already precompiled versions. Requires glslangValidator to be installed on your system." OFF)
+
+# External components
+kompute_option(KOMPUTE_OPT_USE_BUILT_IN_SPDLOG "Use the built-in version of Spdlog. Requires 'KOMPUTE_OPT_USE_SPDLOG' to be set to ON in order to have any effect." ON)
+kompute_option(KOMPUTE_OPT_SPDLOG_ASYNC_MODE "If spdlog is enabled this allows for selecting whether the default logger setup creates sync or async logger" OFF)
+kompute_option(KOMPUTE_OPT_USE_BUILT_IN_FMT "Use the built-in version of fmt." ON)
+kompute_option(KOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER "Use the built-in version of Vulkan Headers. This could be helpful in case your system Vulkan Headers are too new for your driver. If you set this to OFF, please make sure your system Vulkan Headers are supported by your driver." ON)
+kompute_option_string(KOMPUTE_OPT_BUILT_IN_VULKAN_HEADER_TAG "The git tag used for the built-in Vulkan Headers when 'KOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER' is enabled. A list of tags can be found here: https://github.com/KhronosGroup/Vulkan-Headers/tags" "v1.3.231")
+message(STATUS "=======================================================")
+
+# ####################################################
+# Deprecated Options
+# ####################################################
+include(cmake/deprecation_warnings.cmake)
+
+# ####################################################
+# Dependencies
+# ####################################################
+include(cmake/vulkan_shader_compiler.cmake)
+include(cmake/check_vulkan_version.cmake)
+include(FetchContent)
+
+# Vulkan Header
+if(KOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER)
+    FetchContent_Declare(vulkan_header GIT_REPOSITORY https://github.com/KhronosGroup/Vulkan-Headers.git
+        GIT_TAG ${KOMPUTE_OPT_BUILT_IN_VULKAN_HEADER_TAG}) # Source: https://github.com/KhronosGroup/Vulkan-Headers/tags
+    FetchContent_MakeAvailable(vulkan_header)
+
+    if(NOT KOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK)
+        # Ensure the driver supports this Vulkan version
+        check_vulkan_version(INCLUDE_DIR "${vulkan_header_SOURCE_DIR}/include")
+    endif()
+endif()
+
+find_package(Vulkan REQUIRED)
+
+if(Vulkan_FOUND AND NOT TARGET Vulkan::Headers)
+    add_library(Vulkan::Headers INTERFACE IMPORTED)
+    set_target_properties(Vulkan::Headers PROPERTIES
+        INTERFACE_INCLUDE_DIRECTORIES "${Vulkan_INCLUDE_DIRS}")
+endif()
+
+if(NOT KOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER AND NOT KOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK)
+    # Ensure the driver supports this Vulkan version
+    check_vulkan_version(INCLUDE_DIR ${Vulkan_INCLUDE_DIR})
+endif()
+
+# Spdlog
+if(KOMPUTE_OPT_USE_SPDLOG)
+    add_compile_definitions(KOMPUTE_OPT_USE_SPDLOG=1)
+
+    if(NOT KOMPUTE_OPT_LOG_LEVEL_DISABLED)
+        if(KOMPUTE_OPT_USE_BUILT_IN_SPDLOG)
+            set(SPDLOG_BUILD_SHARED ${BUILD_SHARED_LIBS})
+
+            FetchContent_Declare(spdlog GIT_REPOSITORY https://github.com/gabime/spdlog.git
+                GIT_TAG v1.10.0) # Source: https://github.com/gabime/spdlog/releases
+            FetchContent_MakeAvailable(spdlog)
+        else()
+            find_package(spdlog REQUIRED)
+        endif()
+    endif()
+endif()
+
+# fmt
+if(KOMPUTE_OPT_USE_BUILT_IN_FMT)
+    FetchContent_Declare(fmt GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+        GIT_TAG 10.0.0) # Source: https://github.com/fmtlib/fmt/releases
+    FetchContent_MakeAvailable(fmt)
+else()
+    find_package(fmt REQUIRED)
+endif()
+
+add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
+
+# ####################################################
+# Preprocessor Macros
+# ####################################################
+if(KOMPUTE_OPT_DISABLE_VK_DEBUG_LAYERS)
+    add_compile_definitions(KOMPUTE_DISABLE_VK_DEBUG_LAYERS=1)
+endif()
+
+if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
+else()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra -Wpedantic -Werror")
+endif()
+
+# If glslang is cloned, then SPIRV/GlslangToSpv.h will be used instead of glslang/SPIRV/GlslangToSpv.h
+# As after installation, SPIRV/ header files will be found in glslang/SPIRV/ , more info in #193
+if(KOMPUTE_OPT_REPO_SUBMODULE_BUILD)
+    add_definitions(-DUSE_EXTERNAL_GLSLANG)
+endif()
+
+# Allow scripts to call main kompute Makefile
+function(kompute_make KOMPUTE_MAKE_TARGET)
+    add_custom_target(${KOMPUTE_MAKE_TARGET}
+        COMMAND make -C ${PROJECT_SOURCE_DIR} ${KOMPUTE_MAKE_TARGET})
+endfunction()
+
+add_executable(xxd external/bin/xxd.c)
+
+add_subdirectory(src)

kompute/LICENSE

@@ -0,0 +1,203 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2021 The Institute for Ethical AI & Machine Learning
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+

kompute/Makefile

@@ -0,0 +1,210 @@
+# This makefile is optimized to be run from WSL and to interact with the 
+# Windows host as there are limitations when building GPU programs. This
+# makefile contains the commands for interacting with the visual studio
+# build via command line for faster iterations, as the intention is to 
+# support other editors (optimised for vim). There are also commands that
+# support the builds for linux-native compilations and these are the commands
+# starting with mk_.
+
+VERSION := $(shell cat ./VERSION)
+
+VCPKG_WIN_PATH ?= "C:\\Users\\axsau\\Programming\\lib\\vcpkg\\scripts\\buildsystems\\vcpkg.cmake"
+VCPKG_UNIX_PATH ?= "/c/Users/axsau/Programming/lib/vcpkg/scripts/buildsystems/vcpkg.cmake"
+
+# These are the tests that don't work with swiftshader but can be run directly with vulkan
+FILTER_TESTS ?= "-TestAsyncOperations.TestManagerParallelExecution:TestSequence.SequenceTimestamps:TestPushConstants.TestConstantsDouble"
+
+ifeq ($(OS),Windows_NT)     # is Windows_NT on XP, 2000, 7, Vista, 10...
+	CMAKE_BIN ?= "C:\Program Files\CMake\bin\cmake.exe"
+	SCMP_BIN="C:\\VulkanSDK\\1.2.141.2\\Bin32\\glslangValidator.exe"
+	MSBUILD_BIN ?= "C:\\Program Files (x86)\\Microsoft Visual Studio\\2019\\Community\\MSBuild\\Current\\Bin\\MSBuild.exe"
+else
+	CLANG_FORMAT_BIN ?= "/home/alejandro/Programming/lib/clang+llvm-10.0.0-x86_64-linux-gnu-ubuntu-18.04/bin/clang-format"
+	CMAKE_BIN ?= "/c/Program Files/CMake/bin/cmake.exe"
+	MSBUILD_BIN ?= "/c/Program Files (x86)/Microsoft Visual Studio/2019/Community/MSBuild/Current/Bin/MSBuild.exe"
+	# Choosing the binary based on whether it's on WSL or linux-native
+	KERNEL := $(shell uname -r)
+	IS_WSL := $(shell (if [[ "$(KERNEL)" =~ Microsoft$  ]]; then echo '0'; fi))
+	ifeq ($(IS_WSL),0)
+		SCMP_BIN ?= "/c/VulkanSDK/1.2.141.2/Bin32/glslangValidator.exe"
+	else
+		SCMP_BIN ?= "/usr/bin/glslangValidator"
+	endif
+endif
+
+
+####### Main Target Rules #######
+
+push_docs_to_ghpages:
+	GIT_DEPLOY_DIR="build/docs/sphinx/" \
+		GIT_DEPLOY_BRANCH="gh-pages" \
+		GIT_DEPLOY_REPO="origin" \
+			./scripts/push_folder_to_branch.sh
+
+####### CMAKE quickstart commands #######
+
+clean_cmake:
+	rm -rf build/
+
+####### Visual studio build shortcut commands #######
+
+MK_BUILD_TYPE ?= "Release"
+MK_INSTALL_PATH ?= "build/src/CMakeFiles/Export/" # Set to "" if prefer default
+MK_CMAKE_EXTRA_FLAGS ?= ""
+MK_KOMPUTE_EXTRA_CXX_FLAGS ?= ""
+
+mk_cmake:
+	cmake \
+		-Bbuild \
+		-DCMAKE_CXX_FLAGS=$(MK_KOMPUTE_EXTRA_CXX_FLAGS) \
+		-DCMAKE_BUILD_TYPE=$(MK_BUILD_TYPE) \
+		-DCMAKE_INSTALL_PREFIX=$(MK_INSTALL_PATH) \
+		-DKOMPUTE_OPT_INSTALL=ON \
+		-DKOMPUTE_OPT_BUILD_TESTS=ON \
+		-DKOMPUTE_OPT_BUILD_DOCS=ON \
+		-DKOMPUTE_OPT_BUILD_SHADERS=ON \
+		-DKOMPUTE_OPT_CODE_COVERAGE=ON \
+		-DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
+		-DKOMPUTE_OPT_LOG_LEVEL=Debug \
+		$(MK_CMAKE_EXTRA_FLAGS) \
+		-G "Unix Makefiles"
+
+mk_build_all:
+	cmake --build build/. --parallel
+
+mk_build_docs:
+	cmake --build build/. --target gendocsall --parallel
+
+mk_build_kompute:
+	cmake --build build/. --target kompute --parallel
+
+mk_build_tests:
+	cmake --build build/. --target kompute_tests --parallel
+
+mk_run_docs: mk_build_docs
+	(cd build/docs/sphinx && python2.7 -m SimpleHTTPServer)
+
+# An alternative would be: ctest -vv --test-dir build/.
+# But this is not possible since we need to filter specific tests, not complete executables, which is not possible with ctest.
+# https://gitlab.kitware.com/cmake/cmake/-/issues/13168 
+mk_run_tests: mk_build_tests
+	./build/bin/kompute_tests --gtest_filter=$(FILTER_TESTS)
+
+mk_build_swiftshader_library:
+	git clone https://github.com/google/swiftshader || echo "Assuming already cloned"
+	# GCC 8 or above is required otherwise error on "filesystem" lib will appear
+	CC="/usr/bin/gcc-8" CXX="/usr/bin/g++-8" cmake swiftshader/. -Bswiftshader/build/
+	cmake --build swiftshader/build/. --parallel
+
+mk_run_tests_cpu: export VK_ICD_FILENAMES=$(PWD)/swiftshader/build/vk_swiftshader_icd.json
+mk_run_tests_cpu: mk_build_swiftshader_library mk_build_tests mk_run_tests_cpu_only
+
+
+####### Visual studio build shortcut commands #######
+
+VS_BUILD_TYPE ?= "Debug"
+# Run with multiprocessin / parallel build by default
+VS_CMAKE_EXTRA_FLAGS ?= ""
+VS_KOMPUTE_EXTRA_CXX_FLAGS ?= ""
+VS_INSTALL_PATH ?= "build/src/CMakeFiles/Export/" # Set to "" if prefer default
+
+vs_cmake:
+	$(CMAKE_BIN) \
+		-Bbuild \
+		$(VS_CMAKE_EXTRA_FLAGS) \
+		-DCMAKE_TOOLCHAIN_FILE=$(VCPKG_WIN_PATH) \
+		-DCMAKE_CXX_FLAGS=$(VS_KOMPUTE_EXTRA_CXX_FLAGS) \
+		-DCMAKE_INSTALL_PREFIX=$(VS_INSTALL_PATH) \
+		-DKOMPUTE_OPT_INSTALL=ON \
+		-DKOMPUTE_OPT_BUILD_TESTS=ON \
+		-DKOMPUTE_OPT_BUILD_SHADERS=ON \
+		-DKOMPUTE_OPT_CODE_COVERAGE=OFF \
+		-DKOMPUTE_OPT_BUILD_DOCS=OFF \
+		-G "Visual Studio 16 2019" \
+		-DCMAKE_BUILD_TYPE=$(VS_BUILD_TYPE)
+
+vs_build_all:
+	cmake --build build/. --parallel
+
+vs_build_docs:
+	cmake --build build/. --target gendocsall --parallel
+
+vs_install_kompute:
+	cmake --build build/. --target install --parallel
+
+vs_build_kompute:
+	cmake --build build/. --target kompute --parallel
+
+vs_build_tests:
+	cmake --build build/. --target kompute_tests --parallel
+
+vs_run_docs: vs_build_docs
+	(cd build/docs/sphinx && python2.7 -m SimpleHTTPServer)
+
+vs_run_tests: vs_build_tests
+	./build/test/$(VS_BUILD_TYPE)/bin/kompute_tests.exe --gtest_filter=$(FILTER_TESTS)
+
+
+#### PYTHONG ####
+
+test_python:
+	python3 -m pytest -s --log-cli-level=DEBUG -v python/test/
+
+####### Run CI Commands #######
+
+# This command uses act to replicate github action
+# https://github.com/nektos/act
+run_ci:
+	act
+
+####### General project commands #######
+
+generate_python_docstrings:
+	python -m pybind11_mkdoc \
+		-o python/src/docstrings.hpp \
+		kompute/Kompute.hpp \
+		-Iexternal/fmt/include/ \
+		-Iexternal/spdlog/include/ \
+		-Iexternal/glslang/ \
+		-I/usr/include/c++/7.5.0/
+
+install_python_reqs:
+	python3 -m pip install -r scripts/requirements.txt
+
+install_lcov:
+	sudo apt install lcov -y
+
+build_shaders:
+	python3 scripts/convert_shaders.py \
+		--shader-path shaders/glsl \
+		--shader-binary $(SCMP_BIN) \
+		--header-path src/include/kompute/shaders/ \
+		-v
+	python3 scripts/convert_shaders.py \
+		--shader-path test/shaders/glsl \
+		--shader-binary $(SCMP_BIN) \
+		--header-path test/compiled_shaders_include/kompute_test/shaders/ \
+		-v
+
+build_single_header:
+	quom \
+		--include_directory \
+		"src/include/" \
+		"single_include/AggregateHeaders.cpp" \
+		"single_include/kompute/Kompute.hpp"
+
+win_build_xxd:
+	cd external/bin/ && gcc.exe -o xxd.exe xxd.c -DCYGWIN
+
+format:
+	for val in "examples single_include src test" ; do \
+    	find $$val -depth -iname *.h -or -iname *.c -or -iname *.hpp -or -iname *.cpp | grep -v "shaders" | xargs $(CLANG_FORMAT_BIN) -style=file -i; \
+	done
+
+static_scan:
+	cppcheck --project=build/compile_commands.json -iexternal/
+
+build_changelog:
+	docker run --rm -it -v "$(PWD)":/usr/local/src/your-app -e CHANGELOG_GITHUB_TOKEN=${CHANGELOG_GITHUB_TOKEN} ferrarimarco/github-changelog-generator:1.15.2 -u KomputeProject -p kompute
+	chmod 664 CHANGELOG.md # (Read+Write, Read+Write, Read)
+	sed -i -e 's/\(HEAD\|Unreleased\)/v${VERSION}/g' CHANGELOG.md # Replacing unreleased version with latest tag

kompute/README.md

@@ -0,0 +1,513 @@
+
+![GitHub](https://img.shields.io/badge/Version-0.7.0-green.svg)
+![GitHub](https://img.shields.io/badge/C++-14—20-purple.svg)
+![GitHub](https://img.shields.io/badge/Build-cmake-red.svg)
+![GitHub](https://img.shields.io/badge/Python-3.7—3.9-blue.svg)
+![GitHub](https://img.shields.io/badge/License-Apache-black.svg)
+[![CII Best Practices](https://bestpractices.coreinfrastructure.org/projects/4834/badge)](https://bestpractices.coreinfrastructure.org/projects/4834)
+
+<table>
+<tr>
+
+<td width="20%">
+<img src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/kompute.jpg">
+</td>
+
+<td>
+
+<h1>Kompute</h1>
+<h3>The general purpose GPU compute framework for cross vendor graphics cards (AMD, Qualcomm, NVIDIA & friends)</h3>
+
+</td>
+
+</tr>
+</table>
+
+<h4>Blazing fast, mobile-enabled, asynchronous, and optimized for advanced GPU acceleration usecases.</h4>
+
+💬 [Join the Discord & Community Calls](https://kompute.cc/overview/community.html) 🔋 [Documentation](https://kompute.cc) 💻 [Blog Post](https://medium.com/@AxSaucedo/machine-learning-and-data-processing-in-the-gpu-with-vulkan-kompute-c9350e5e5d3a) ⌨ [Examples](#more-examples) 💾
+
+<hr>
+
+##### Kompute is backed by the Linux Foundation as a <a href="https://lfaidata.foundation/blog/2021/08/26/kompute-joins-lf-ai-data-as-new-sandbox-project/">hosted project</a> by the LF AI & Data Foundation.
+
+<table>
+<tr>
+<td>
+<a href="https://www.linuxfoundation.org/projects/">
+<img src="https://upload.wikimedia.org/wikipedia/commons/b/b5/Linux_Foundation_logo.png">
+</a>
+</td>
+<td>
+<a href="https://lfaidata.foundation/projects/">
+<img src="https://raw.githubusercontent.com/lfai/artwork/main/lfaidata-assets/lfaidata/horizontal/color/lfaidata-horizontal-color.png">
+</a>
+</td>
+</tr>
+</table>
+
+
+## Principles & Features
+
+* [Flexible Python module](#your-first-kompute-python) with [C++ SDK](#your-first-kompute-c) for optimizations
+* [Asynchronous & parallel processing](#asynchronous-and-parallel-operations) support through GPU family queues
+* [Mobile enabled](#mobile-enabled) with examples via Android NDK across several architectures
+* BYOV: [Bring-your-own-Vulkan design](#motivations) to play nice with existing Vulkan applications
+* Explicit relationships for GPU and host [memory ownership and memory management](https://kompute.cc/overview/memory-management.html)
+* Robust codebase with [90% unit test code coverage](https://kompute.cc/codecov/)
+* Advanced use-cases on [machine learning 🤖](https://towardsdatascience.com/machine-learning-and-data-processing-in-the-gpu-with-vulkan-kompute-c9350e5e5d3a), [mobile development 📱](https://towardsdatascience.com/gpu-accelerated-machine-learning-in-your-mobile-applications-using-the-android-ndk-vulkan-kompute-1e9da37b7617) and [game development 🎮](https://towardsdatascience.com/supercharging-game-development-with-gpu-accelerated-ml-using-vulkan-kompute-the-godot-game-engine-4e75a84ea9f0).
+* Active community with [monthly calls, discord chat and more](https://kompute.cc/overview/community.html)
+
+![](https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/komputer-logos.gif)
+
+## Getting Started
+
+Below you can find a GPU multiplication example using the C++ and Python Kompute interfaces.
+
+You can [join the Discord](https://discord.gg/MaH5Jv5zwv) for questions / discussion, open a [github issue](https://github.com/KomputeProject/kompute/issues/new), or read [the documentation](https://kompute.cc/).
+
+### Your First Kompute (C++)
+
+The C++ interface provides low level access to the native components of Kompute, enabling for [advanced optimizations](https://kompute.cc/overview/async-parallel.html) as well as [extension of components](https://kompute.cc/overview/reference.html).
+
+```c++
+
+void kompute(const std::string& shader) {
+
+    // 1. Create Kompute Manager with default settings (device 0, first queue and no extensions)
+    kp::Manager mgr; 
+
+    // 2. Create and initialise Kompute Tensors through manager
+
+    // Default tensor constructor simplifies creation of float values
+    auto tensorInA = mgr.tensor({ 2., 2., 2. });
+    auto tensorInB = mgr.tensor({ 1., 2., 3. });
+    // Explicit type constructor supports uint32, int32, double, float and bool
+    auto tensorOutA = mgr.tensorT<uint32_t>({ 0, 0, 0 });
+    auto tensorOutB = mgr.tensorT<uint32_t>({ 0, 0, 0 });
+
+    std::vector<std::shared_ptr<kp::Tensor>> params = {tensorInA, tensorInB, tensorOutA, tensorOutB};
+
+    // 3. Create algorithm based on shader (supports buffers & push/spec constants)
+    kp::Workgroup workgroup({3, 1, 1});
+    std::vector<float> specConsts({ 2 });
+    std::vector<float> pushConstsA({ 2.0 });
+    std::vector<float> pushConstsB({ 3.0 });
+
+    auto algorithm = mgr.algorithm(params,
+                                   // See documentation shader section for compileSource
+                                   compileSource(shader),
+                                   workgroup,
+                                   specConsts,
+                                   pushConstsA);
+
+    // 4. Run operation synchronously using sequence
+    mgr.sequence()
+        ->record<kp::OpTensorSyncDevice>(params)
+        ->record<kp::OpAlgoDispatch>(algorithm) // Binds default push consts
+        ->eval() // Evaluates the two recorded operations
+        ->record<kp::OpAlgoDispatch>(algorithm, pushConstsB) // Overrides push consts
+        ->eval(); // Evaluates only last recorded operation
+
+    // 5. Sync results from the GPU asynchronously
+    auto sq = mgr.sequence();
+    sq->evalAsync<kp::OpTensorSyncLocal>(params);
+
+    // ... Do other work asynchronously whilst GPU finishes
+
+    sq->evalAwait();
+
+    // Prints the first output which is: { 4, 8, 12 }
+    for (const float& elem : tensorOutA->vector()) std::cout << elem << "  ";
+    // Prints the second output which is: { 10, 10, 10 }
+    for (const float& elem : tensorOutB->vector()) std::cout << elem << "  ";
+
+} // Manages / releases all CPU and GPU memory resources
+
+int main() {
+
+    // Define a raw string shader (or use the Kompute tools to compile to SPIRV / C++ header
+    // files). This shader shows some of the main components including constants, buffers, etc
+    std::string shader = (R"(
+        #version 450
+
+        layout (local_size_x = 1) in;
+
+        // The input tensors bind index is relative to index in parameter passed
+        layout(set = 0, binding = 0) buffer buf_in_a { float in_a[]; };
+        layout(set = 0, binding = 1) buffer buf_in_b { float in_b[]; };
+        layout(set = 0, binding = 2) buffer buf_out_a { uint out_a[]; };
+        layout(set = 0, binding = 3) buffer buf_out_b { uint out_b[]; };
+
+        // Kompute supports push constants updated on dispatch
+        layout(push_constant) uniform PushConstants {
+            float val;
+        } push_const;
+
+        // Kompute also supports spec constants on initalization
+        layout(constant_id = 0) const float const_one = 0;
+
+        void main() {
+            uint index = gl_GlobalInvocationID.x;
+            out_a[index] += uint( in_a[index] * in_b[index] );
+            out_b[index] += uint( const_one * push_const.val );
+        }
+    )");
+
+    // Run the function declared above with our raw string shader
+    kompute(shader);
+}
+
+```
+
+### Your First Kompute (Python)
+
+The [Python package](https://kompute.cc/overview/python-package.html) provides a [high level interactive interface](https://kompute.cc/overview/python-reference.html) that enables for experimentation whilst ensuring high performance and fast development workflows.
+
+```python
+
+from .utils import compile_source # using util function from python/test/utils
+
+def kompute(shader):
+    # 1. Create Kompute Manager with default settings (device 0, first queue and no extensions)
+    mgr = kp.Manager()
+
+    # 2. Create and initialise Kompute Tensors through manager
+
+    # Default tensor constructor simplifies creation of float values
+    tensor_in_a = mgr.tensor([2, 2, 2])
+    tensor_in_b = mgr.tensor([1, 2, 3])
+    # Explicit type constructor supports uint32, int32, double, float and bool
+    tensor_out_a = mgr.tensor_t(np.array([0, 0, 0], dtype=np.uint32))
+    tensor_out_b = mgr.tensor_t(np.array([0, 0, 0], dtype=np.uint32))
+
+    params = [tensor_in_a, tensor_in_b, tensor_out_a, tensor_out_b]
+
+    # 3. Create algorithm based on shader (supports buffers & push/spec constants)
+    workgroup = (3, 1, 1)
+    spec_consts = [2]
+    push_consts_a = [2]
+    push_consts_b = [3]
+
+    # See documentation shader section for compile_source
+    spirv = compile_source(shader)
+
+    algo = mgr.algorithm(params, spirv, workgroup, spec_consts, push_consts_a)
+
+    # 4. Run operation synchronously using sequence
+    (mgr.sequence()
+        .record(kp.OpTensorSyncDevice(params))
+        .record(kp.OpAlgoDispatch(algo)) # Binds default push consts provided
+        .eval() # evaluates the two recorded ops
+        .record(kp.OpAlgoDispatch(algo, push_consts_b)) # Overrides push consts
+        .eval()) # evaluates only the last recorded op
+
+    # 5. Sync results from the GPU asynchronously
+    sq = mgr.sequence()
+    sq.eval_async(kp.OpTensorSyncLocal(params))
+
+    # ... Do other work asynchronously whilst GPU finishes
+
+    sq.eval_await()
+
+    # Prints the first output which is: { 4, 8, 12 }
+    print(tensor_out_a)
+    # Prints the first output which is: { 10, 10, 10 }
+    print(tensor_out_b)
+
+if __name__ == "__main__":
+
+    # Define a raw string shader (or use the Kompute tools to compile to SPIRV / C++ header
+    # files). This shader shows some of the main components including constants, buffers, etc
+    shader = """
+        #version 450
+
+        layout (local_size_x = 1) in;
+
+        // The input tensors bind index is relative to index in parameter passed
+        layout(set = 0, binding = 0) buffer buf_in_a { float in_a[]; };
+        layout(set = 0, binding = 1) buffer buf_in_b { float in_b[]; };
+        layout(set = 0, binding = 2) buffer buf_out_a { uint out_a[]; };
+        layout(set = 0, binding = 3) buffer buf_out_b { uint out_b[]; };
+
+        // Kompute supports push constants updated on dispatch
+        layout(push_constant) uniform PushConstants {
+            float val;
+        } push_const;
+
+        // Kompute also supports spec constants on initalization
+        layout(constant_id = 0) const float const_one = 0;
+
+        void main() {
+            uint index = gl_GlobalInvocationID.x;
+            out_a[index] += uint( in_a[index] * in_b[index] );
+            out_b[index] += uint( const_one * push_const.val );
+        }
+    """
+
+    kompute(shader)
+
+```
+
+### Interactive Notebooks & Hands on Videos
+
+You are able to try out the interactive Colab Notebooks which allow you to use a free GPU. The available examples are the Python and C++ examples below:
+
+<table>
+<tr>
+
+<td width="50%">
+<h5>Try the interactive <a href="https://colab.research.google.com/drive/1l3hNSq2AcJ5j2E3YIw__jKy5n6M615GP?usp=sharing">C++ Colab</a> from <a href="https://towardsdatascience.com/machine-learning-and-data-processing-in-the-gpu-with-vulkan-kompute-c9350e5e5d3a">Blog Post</a></h5>
+</td>
+
+<td>
+<h5>Try the interactive <a href="https://colab.research.google.com/drive/15uQ7qMZuOyk8JcXF-3SB2R5yNFW21I4P">Python Colab</a> from <a href="https://towardsdatascience.com/beyond-cuda-gpu-accelerated-python-for-machine-learning-in-cross-vendor-graphics-cards-made-simple-6cc828a45cc3">Blog Post</a></h5>
+</td>
+
+</tr>
+<tr>
+
+<td width="50%">
+<a href="https://colab.research.google.com/drive/1l3hNSq2AcJ5j2E3YIw__jKy5n6M615GP?authuser=1#scrollTo=1BipBsO-fQRD">
+<img src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/binder-cpp.jpg">
+</a>
+</td>
+
+<td>
+<a href="https://colab.research.google.com/drive/15uQ7qMZuOyk8JcXF-3SB2R5yNFW21I4P">
+<img src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/binder-python.jpg">
+</a>
+</td>
+
+</tr>
+</table>
+
+
+You can also check out the two following talks presented at the FOSDEM 2021 conference. 
+
+Both videos have timestamps which will allow you to skip to the most relevant section for you - the intro & motivations for both is almost the same so you can skip to the more specific content.
+
+<table>
+<tr>
+
+<td width="50%">
+<h5>Watch the video for <a href="https://www.youtube.com/watch?v=Xz4fiQNmGSA">C++ Enthusiasts</a> </h5>
+</td>
+
+<td>
+<h5>Watch the video for <a href="https://www.youtube.com/watch?v=AJRyZ09IUdg">Python & Machine Learning</a> Enthusiasts</h5>
+</td>
+
+</tr>
+<tr>
+
+<td width="50%">
+<a href="https://www.youtube.com/watch?v=Xz4fiQNmGSA">
+<img src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/kompute-cpp-video.png">
+</a>
+</td>
+
+<td>
+<a href="https://www.youtube.com/watch?v=AJRyZ09IUdg">
+<img src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/kompute-python-video.png">
+</a>
+</td>
+
+</tr>
+</table>
+
+
+## Architectural Overview
+
+The core architecture of Kompute includes the following:
+* [Kompute Manager](https://kompute.cc/overview/reference.html#manager) - Base orchestrator which creates and manages device and child components
+* [Kompute Sequence](https://kompute.cc/overview/reference.html#sequence) - Container of operations that can be sent to GPU as batch
+* [Kompute Operation (Base)](https://kompute.cc/overview/reference.html#algorithm) - Base class from which all operations inherit
+* [Kompute Tensor](https://kompute.cc/overview/reference.html#tensor) - Tensor structured data used in GPU operations
+* [Kompute Algorithm](https://kompute.cc/overview/reference.html#algorithm) - Abstraction for (shader) logic executed in the GPU
+
+To see a full breakdown you can read further in the [C++ Class Reference](https://kompute.cc/overview/reference.html).
+
+<table>
+<th>
+Full Architecture
+</th>
+<th>
+Simplified Kompute Components
+</th>
+<tr>
+<td width=30%>
+
+
+<img width="100%" src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/kompute-vulkan-architecture.jpg">
+
+<br>
+<br>
+(very tiny, check the <a href="https://ethicalml.github.io/vulkan-kompute/overview/reference.html">full reference diagram in docs for details</a>)
+<br>
+<br>
+
+<img width="100%" src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/suspicious.jfif">
+
+</td>
+<td>
+<img width="100%" src="https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/kompute-architecture.jpg">
+</td>
+</tr>
+</table>
+
+
+## Asynchronous and Parallel Operations
+
+Kompute provides flexibility to run operations in an asynrchonous way through vk::Fences. Furthermore, Kompute enables for explicit allocation of queues, which allow for parallel execution of operations across queue families.
+
+The image below provides an intuition on how Kompute Sequences can be allocated to different queues to enable parallel execution based on hardware. You can see the [hands on example](https://kompute.cc/overview/advanced-examples.html#parallel-operations), as well as the [detailed documentation page](https://kompute.cc/overview/async-parallel.html) describing how it would work using an NVIDIA 1650 as an example. 
+
+![](https://raw.githubusercontent.com/KomputeProject/kompute/master/docs/images/queue-allocation.jpg)
+
+## Mobile Enabled
+
+Kompute has been optimized to work in mobile environments. The [build system](#build-overview) enables for dynamic loading of the Vulkan shared library for Android environments, together with a working [Android NDK wrapper](https://github.com/KomputeProject/kompute/tree/master/vk_ndk_wrapper_include) for the CPP headers.
+
+<table>
+<tr>
+
+<td width="70%">
+<p>
+For a full deep dive you can read the blog post "<a href="https://towardsdatascience.com/gpu-accelerated-machine-learning-in-your-mobile-applications-using-the-android-ndk-vulkan-kompute-1e9da37b7617">Supercharging your Mobile Apps with On-Device GPU Accelerated Machine Learning</a>". 
+
+You can also access the <a href="https://github.com/KomputeProject/kompute/tree/v0.4.0/examples/android/android-simple">end-to-end example code</a> in the repository, which can be run using android studio.
+
+</p>
+
+
+<img src="https://raw.githubusercontent.com/KomputeProject/kompute/android-example/docs/images/android-editor.jpg">
+
+</td>
+
+
+<td width="30%">
+<img src="https://raw.githubusercontent.com/KomputeProject/kompute/android-example/docs/images/android-kompute.jpg">
+</td>
+
+</tr>
+</table>
+
+## More examples
+
+### Simple examples
+
+* [Simple multiplication example](https://kompute.cc/overview/advanced-examples.html#simple-shader-example)
+* [Record batch commands with a Kompute Sequence](https://kompute.cc/overview/advanced-examples.html#record-batch-commands)
+* [Run Asynchronous Operations](https://kompute.cc/overview/advanced-examples.html#asynchronous-operations)
+* [Run Parallel Operations Across Multiple GPU Queues](https://kompute.cc/overview/advanced-examples.html#parallel-operations)
+* [Create your custom Kompute Operations](https://kompute.cc/overview/advanced-examples.html#your-custom-kompute-operation)
+* [Implementing logistic regression from scratch](https://kompute.cc/overview/advanced-examples.html#logistic-regression-example)
+
+### End-to-end examples
+
+* [Machine Learning Logistic Regression Implementation](https://towardsdatascience.com/machine-learning-and-data-processing-in-the-gpu-with-vulkan-kompute-c9350e5e5d3a)
+* [Parallelizing GPU-intensive Workloads via Multi-Queue Operations](https://towardsdatascience.com/parallelizing-heavy-gpu-workloads-via-multi-queue-operations-50a38b15a1dc)
+* [Android NDK Mobile Kompute ML Application](https://towardsdatascience.com/gpu-accelerated-machine-learning-in-your-mobile-applications-using-the-android-ndk-vulkan-kompute-1e9da37b7617)
+* [Game Development Kompute ML in Godot Engine](https://towardsdatascience.com/supercharging-game-development-with-gpu-accelerated-ml-using-vulkan-kompute-the-godot-game-engine-4e75a84ea9f0)
+
+## Python Package
+
+Besides the C++ core SDK you can also use the Python package of Kompute, which exposes the same core functionality, and supports interoperability with Python objects like Lists, Numpy Arrays, etc.
+
+The only dependencies are Python 3.5+ and Cmake 3.4.1+. You can install Kompute from the [Python pypi package](https://pypi.org/project/kp/) using the following command.
+
+```
+pip install kp
+```
+
+You can also install from master branch using:
+
+```
+pip install git+git://github.com/KomputeProject/kompute.git@master
+```
+
+For further details you can read the [Python Package documentation](https://kompute.cc/overview/python-package.html) or the [Python Class Reference documentation](https://kompute.cc/overview/python-reference.html).
+
+## C++ Build Overview
+
+The build system provided uses `cmake`, which allows for cross platform builds.
+
+The top level `Makefile` provides a set of optimized configurations for development as well as the docker image build, but you can start a build with the following command:
+
+```
+   cmake -Bbuild
+```
+
+You also are able to add Kompute in your repo with `add_subdirectory` - the [Android example CMakeLists.txt file](https://github.com/KomputeProject/kompute/blob/7c8c0eeba2cdc098349fcd999102bb2cca1bf711/examples/android/android-simple/app/src/main/cpp/CMakeLists.txt#L3) shows how this would be done.
+
+For a more advanced overview of the build configuration check out the [Build System Deep Dive](https://kompute.cc/overview/build-system.html) documentation.
+
+## Kompute Development
+
+We appreciate PRs and Issues. If you want to contribute try checking the "Good first issue" tag, but even using Kompute and reporting issues is a great contribution!
+
+### Contributing
+
+#### Dev Dependencies
+
+* Testing
+    + GTest
+* Documentation
+    + Doxygen (with Dot)
+    + Sphynx
+
+#### Development
+
+* Follows Mozilla C++ Style Guide https://www-archive.mozilla.org/hacking/mozilla-style-guide.html
+    + Uses post-commit hook to run the linter, you can set it up so it runs the linter before commit
+    + All dependencies are defined in vcpkg.json 
+* Uses cmake as build system, and provides a top level makefile with recommended command
+* Uses xxd (or xxd.exe windows 64bit port) to convert shader spirv to header files
+* Uses doxygen and sphinx for documentation and autodocs
+* Uses vcpkg for finding the dependencies, it's the recommended set up to retrieve the libraries
+
+If you want to run with debug layers you can add them with the `KOMPUTE_ENV_DEBUG_LAYERS` parameter as:
+
+```
+export KOMPUTE_ENV_DEBUG_LAYERS="VK_LAYER_LUNARG_api_dump"
+```
+
+##### Updating documentation
+
+To update the documentation you will need to:
+* Run the gendoxygen target in the build system
+* Run the gensphynx target in the build-system 
+* Push to github pages with `make push_docs_to_ghpages`
+
+##### Running tests
+
+Running the unit tests has been significantly simplified for contributors.
+
+The tests run on CPU, and can be triggered using the ACT command line interface (https://github.com/nektos/act) - once you install the command line (And start the Docker daemon) you just have to type:
+
+```
+$ act
+
+[Python Tests/python-tests] 🚀  Start image=axsauze/kompute-builder:0.2
+[C++ Tests/cpp-tests      ] 🚀  Start image=axsauze/kompute-builder:0.2
+[C++ Tests/cpp-tests      ]   🐳  docker run image=axsauze/kompute-builder:0.2 entrypoint=["/usr/bin/tail" "-f" "/dev/null"] cmd=[]
+[Python Tests/python-tests]   🐳  docker run image=axsauze/kompute-builder:0.2 entrypoint=["/usr/bin/tail" "-f" "/dev/null"] cmd=[]
+...
+```
+
+The repository contains unit tests for the C++ and Python code, and can be found under the `test/` and `python/test` folder.
+
+The tests are currently run through the CI using Github Actions. It uses the images found in `docker-builders/`.
+
+In order to minimise hardware requirements the tests can run without a GPU, directly in the CPU using [Swiftshader](https://github.com/google/swiftshader).
+
+For more information on how the CI and tests are setup, you can go to the [CI, Docker and Tests Section](https://kompute.cc/overview/ci-tests.html) in the documentation.
+
+## Motivations
+
+This project started after seeing that a lot of new and renowned ML & DL projects like Pytorch, Tensorflow, Alibaba DNN, Tencent NCNN - among others - have either integrated or are looking to integrate the Vulkan SDK to add mobile (and cross-vendor) GPU support.
+
+The Vulkan SDK offers a great low level interface that enables for highly specialized optimizations - however it comes at a cost of highly verbose code which requires 500-2000 lines of code to even begin writing application code. This has resulted in each of these projects having to implement the same baseline to abstract the non-compute related features of the Vulkan SDK. This large amount of non-standardised boiler-plate can result in limited knowledge transfer, higher chance of unique framework implementation bugs being introduced, etc.
+
+We are currently developing Kompute not to hide the Vulkan SDK interface (as it's incredibly well designed) but to augment it with a direct focus on the Vulkan SDK's GPU computing capabilities. [This article](https://towardsdatascience.com/machine-learning-and-data-processing-in-the-gpu-with-vulkan-kompute-c9350e5e5d3a) provides a high level overview of the motivations of Kompute, together with a set of hands on examples that introduce both GPU computing as well as the core Kompute architecture.

kompute/cmake/bin2h.cmake

@@ -0,0 +1,106 @@
+##################################################################################
+# Based on: https://github.com/sivachandran/cmake-bin2h
+#
+# Copyright 2020 Sivachandran Paramasivam
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+##################################################################################
+
+include(CMakeParseArguments)
+
+# Function to wrap a given string into multiple lines at the given column position.
+# Parameters:
+#   VARIABLE    - The name of the CMake variable holding the string.
+#   AT_COLUMN   - The column position at which string will be wrapped.
+function(WRAP_STRING)
+    set(oneValueArgs VARIABLE AT_COLUMN)
+    cmake_parse_arguments(WRAP_STRING "${options}" "${oneValueArgs}" "" ${ARGN})
+
+    string(LENGTH ${${WRAP_STRING_VARIABLE}} stringLength)
+    math(EXPR offset "0")
+
+    while(stringLength GREATER 0)
+
+        if(stringLength GREATER ${WRAP_STRING_AT_COLUMN})
+            math(EXPR length "${WRAP_STRING_AT_COLUMN}")
+        else()
+            math(EXPR length "${stringLength}")
+        endif()
+
+        string(SUBSTRING ${${WRAP_STRING_VARIABLE}} ${offset} ${length} line)
+        set(lines "${lines}\n${line}")
+
+        math(EXPR stringLength "${stringLength} - ${length}")
+        math(EXPR offset "${offset} + ${length}")
+    endwhile()
+
+    set(${WRAP_STRING_VARIABLE} "${lines}" PARENT_SCOPE)
+endfunction()
+
+# Function to embed contents of a file as byte array in C/C++ header file(.h). The header file
+# will contain a byte array and integer variable holding the size of the array.
+# Parameters
+#   SOURCE_FILE      - The path of source file whose contents will be embedded in the header file.
+#   VARIABLE_NAME    - The name of the variable for the byte array. The string "_SIZE" will be append
+#                      to this name and will be used a variable name for size variable.
+#   HEADER_FILE      - The path of header file.
+#   APPEND           - If specified appends to the header file instead of overwriting it
+#   NULL_TERMINATE   - If specified a null byte(zero) will be append to the byte array. This will be
+#                      useful if the source file is a text file and we want to use the file contents
+#                      as string. But the size variable holds size of the byte array without this
+#                      null byte.
+#   HEADER_NAMESPACE - The namespace, where the array should be located in.
+#   IS_BIG_ENDIAN    - If set to true, will not revers the byte order for the uint32_t to match the
+#                      big endian system architecture
+# Usage:
+#   bin2h(SOURCE_FILE "Logo.png" HEADER_FILE "Logo.h" VARIABLE_NAME "LOGO_PNG")
+function(BIN2H)
+    set(options APPEND NULL_TERMINATE)
+    set(oneValueArgs SOURCE_FILE VARIABLE_NAME HEADER_FILE)
+    cmake_parse_arguments(BIN2H "${options}" "${oneValueArgs}" "" ${ARGN})
+
+    # reads source file contents as hex string
+    file(READ ${BIN2H_SOURCE_FILE} hexString HEX)
+    string(LENGTH ${hexString} hexStringLength)
+
+    # appends null byte if asked
+    if(BIN2H_NULL_TERMINATE)
+        set(hexString "${hexString}00")
+    endif()
+
+    # wraps the hex string into multiple lines at column 32(i.e. 16 bytes per line)
+    wrap_string(VARIABLE hexString AT_COLUMN 32)
+    math(EXPR arraySize "${hexStringLength} / 8")
+
+    # adds '0x' prefix and comma suffix before and after every byte respectively
+    if(IS_BIG_ENDIAN)
+        message(STATUS "Interpreting shader in big endian...")
+        string(REGEX REPLACE "([0-9a-f][0-9a-f])([0-9a-f][0-9a-f])([0-9a-f][0-9a-f])([0-9a-f][0-9a-f])" "0x\\1\\2\\3\\4, " arrayValues ${hexString})
+    else()
+        message(STATUS "Interpreting shader in little endian...")
+        string(REGEX REPLACE "([0-9a-f][0-9a-f])([0-9a-f][0-9a-f])([0-9a-f][0-9a-f])([0-9a-f][0-9a-f])" "0x\\4\\3\\2\\1, " arrayValues ${hexString})
+    endif()
+    # removes trailing comma
+    string(REGEX REPLACE ", $" "" arrayValues ${arrayValues})
+
+    # converts the variable name into proper C identifier
+    string(MAKE_C_IDENTIFIER "${BIN2H_VARIABLE_NAME}" BIN2H_VARIABLE_NAME)
+    string(TOUPPER "${BIN2H_VARIABLE_NAME}" BIN2H_VARIABLE_NAME)
+
+    # declares byte array and the length variables
+    set(namespaceStart "namespace ${HEADER_NAMESPACE} {")
+    set(namespaceEnd "} // namespace ${HEADER_NAMESPACE}")
+    set(arrayIncludes "#pragma once\n#include <array>\n#include <cstdint>")
+    set(arrayDefinition "const std::array<uint32_t, ${arraySize}> ${BIN2H_VARIABLE_NAME} = { ${arrayValues} };")
+
+    set(declarations "${arrayIncludes}\n\n${namespaceStart}\n${arrayDefinition}\n${namespaceEnd}\n\n")
+    if(BIN2H_APPEND)
+        file(APPEND ${BIN2H_HEADER_FILE} "${declarations}")
+    else()
+        file(WRITE ${BIN2H_HEADER_FILE} "${declarations}")
+    endif()
+endfunction()

kompute/cmake/bin_file_to_header.cmake

@@ -0,0 +1,19 @@
+cmake_minimum_required(VERSION 3.20)
+
+if(${INPUT_SHADER_FILE} STREQUAL "")
+    message(FATAL_ERROR "No input file path provided via 'INPUT_SHADER_FILE'.")
+endif()
+
+if(${OUTPUT_HEADER_FILE} STREQUAL "")
+    message(FATAL_ERROR "No output file path provided via 'OUTPUT_HEADER_FILE'.")
+endif()
+
+if(${HEADER_NAMESPACE} STREQUAL "")
+    message(FATAL_ERROR "No header namespace provided via 'HEADER_NAMESPACE'.")
+endif()
+
+include(bin2h.cmake)
+
+get_filename_component(BINARY_FILE_CONTENT ${INPUT_SHADER_FILE} NAME)
+bin2h(SOURCE_FILE ${INPUT_SHADER_FILE} HEADER_FILE ${OUTPUT_HEADER_FILE} VARIABLE_NAME ${BINARY_FILE_CONTENT} HEADER_NAMESPACE ${HEADER_NAMESPACE})
+file(APPEND ${OUTPUT_HEADER_FILE} "\n")

kompute/cmake/check_vulkan_version.cmake

@@ -0,0 +1,139 @@
+# Current issue: Only checks the result of GPU0
+function(check_vulkan_version)
+    cmake_parse_arguments(VULKAN_CHECK_VERSION "" "INCLUDE_DIR" "" ${ARGN})
+    message(STATUS "Ensuring the currently installed driver supports the Vulkan version requested by the Vulkan Header.")
+
+    # Get the current Vulkan Header version (e.g. 1.2.189).
+    # This snippet is based on: https://gitlab.kitware.com/cmake/cmake/-/blob/v3.23.1/Modules/FindVulkan.cmake#L140-156
+    if(VULKAN_CHECK_VERSION_INCLUDE_DIR)
+        set(VULKAN_CORE_H ${VULKAN_CHECK_VERSION_INCLUDE_DIR}/vulkan/vulkan_core.h)
+        if(EXISTS ${VULKAN_CORE_H})
+            file(STRINGS ${VULKAN_CORE_H} VULKAN_HEADER_VERSION_LINE REGEX "^#define VK_HEADER_VERSION ")
+            string(REGEX MATCHALL "[0-9]+" VULKAN_HEADER_VERSION "${VULKAN_HEADER_VERSION_LINE}")
+            file(STRINGS ${VULKAN_CORE_H} VULKAN_HEADER_VERSION_LINE2 REGEX "^#define VK_HEADER_VERSION_COMPLETE ")
+            if(NOT ${VULKAN_HEADER_VERSION_LINE2} STREQUAL "")
+                string(REGEX MATCHALL "[0-9]+" VULKAN_HEADER_VERSION2 "${VULKAN_HEADER_VERSION_LINE2}")
+                list(LENGTH VULKAN_HEADER_VERSION2 _len)
+                # Versions >= 1.2.175 have an additional numbers in front of e.g. '0, 1, 2' instead of '1, 2'
+                if(_len EQUAL 3)
+                    list(REMOVE_AT VULKAN_HEADER_VERSION2 0)
+                endif()
+                list(APPEND VULKAN_HEADER_VERSION2 ${VULKAN_HEADER_VERSION})
+                list(JOIN VULKAN_HEADER_VERSION2 "." VULKAN_HEADER_VERSION)
+            else()
+                file(STRINGS ${VULKAN_CORE_H} VULKAN_HEADER_API_VERSION_1_2 REGEX "^#define VK_API_VERSION_1_2.*")
+                if(NOT ${VULKAN_HEADER_API_VERSION_1_2} STREQUAL "")
+                    set(VULKAN_HEADER_VERSION "1.2.${VULKAN_HEADER_VERSION}")
+                else()
+                    file(STRINGS ${VULKAN_CORE_H} VULKAN_HEADER_API_VERSION_1_1 REGEX "^#define VK_API_VERSION_1_1.*")
+                    if(NOT ${VULKAN_HEADER_API_VERSION_1_1} STREQUAL "")
+                        set(VULKAN_HEADER_VERSION "1.1.${VULKAN_HEADER_VERSION}")
+                    else()
+                        message(FATAL_ERROR "'${VULKAN_CORE_H}' does not contain a supported Vulkan version. Probably because its < 1.2.0.")
+                    endif()
+                endif()
+            endif()
+        else()
+            message(FATAL_ERROR "'${VULKAN_CORE_H}' does not exist. Try calling 'find_package(Vulkan REQUIRED)' before you call this function or set 'Vulkan_INCLUDE_DIR' manually!")
+            return()
+        endif()
+    else()
+        message(FATAL_ERROR "Invalid Vulkan include directory given. Try calling 'find_package(Vulkan REQUIRED)' before you call this function or set 'Vulkan_INCLUDE_DIR' manually!")
+        return()
+    endif()
+    message(STATUS "Found Vulkan Header version: ${VULKAN_HEADER_VERSION}")
+
+    # Get Vulkan version supported by driver
+    find_program(VULKAN_INFO_PATH NAMES vulkaninfo)
+    if(VULKAN_INFO_PATH STREQUAL "VULKAN_INFO_PATH-NOTFOUND")
+        message(FATAL_ERROR "vulkaninfo not found. The Vulkan SDK might not be installed properly. If you know what you are doing, you can disable the Vulkan version check by setting 'KOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK' to 'ON' (-DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON).")
+        return()
+    endif()
+
+    execute_process(COMMAND "vulkaninfo"
+                    OUTPUT_VARIABLE VULKAN_INFO_OUTPUT
+                    RESULT_VARIABLE VULKAN_INFO_RETURN)
+    if(NOT ${VULKAN_INFO_RETURN} EQUAL 0)
+        message(FATAL_ERROR "Running vulkaninfo failed with return code ${VULKAN_INFO_RETURN}. Make sure you have 'vulkan-tools' installed. Result:\n${VULKAN_INFO_OUTPUT}?")
+        return()
+    else()
+        message(STATUS "Running vulkaninfo was successful. Parsing the output...")
+    endif()
+
+    # Check if running vulkaninfo was successfully
+    string(FIND "${VULKAN_INFO_OUTPUT}" "Vulkan Instance Version" VULKAN_INFO_SUCCESSFUL)
+    if(VULKAN_INFO_SUCCESSFUL LESS 0)
+        message(FATAL_ERROR "Running vulkaninfo failed. Make sure you have 'vulkan-tools' installed and DISPLAY is configured. If you know what you are doing, you can disable the Vulkan version check by setting 'KOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK' to 'ON' (-DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON). Result:\n${VULKAN_INFO_OUTPUT}?")
+    endif()
+
+    string(REGEX MATCHALL "(GPU[0-9]+)" GPU_IDS "${VULKAN_INFO_OUTPUT}")
+    if(NOT GPU_IDS)
+        message(FATAL_ERROR "No GPU supporting Vulkan found in vulkaninfo. Does your GPU (driver) support Vulkan?")
+    endif()
+
+    string(REGEX MATCHALL "apiVersion[ ]*=[ ]*[0-9a-fA-F]*[ ]*[(]*([0-9]+[.][0-9]+[.][0-9]+)[)]*" GPU_API_VERSIONS ${VULKAN_INFO_OUTPUT})
+    if(NOT GPU_API_VERSIONS)
+        message(FATAL_ERROR "No valid Vulkan API version found in vulkaninfo. Does your GPU (driver) support Vulkan?")
+    endif()
+
+    # Check length
+    # message(FATAL_ERROR "GPUS: ${GPU_IDS}")
+    list(LENGTH GPU_IDS GPU_IDS_LENGTH)
+    list(LENGTH GPU_API_VERSIONS GPU_API_VERSIONS_LENGTH)
+    if(NOT ${GPU_IDS_LENGTH} EQUAL ${GPU_API_VERSIONS_LENGTH})
+        message(FATAL_ERROR "Found ${GPU_IDS_LENGTH} GPUs, but ${GPU_API_VERSIONS_LENGTH} API versions in vulkaninfo. We expected to find an equal amount of them.")
+    endif()
+
+    # Compare versions
+    set(VALID_GPU "")
+    set(VALID_VULKAN_VERSION "")
+    math(EXPR ITER_LEN "${GPU_IDS_LENGTH} - 1")
+    foreach(INDEX RANGE ${ITER_LEN})
+        list(GET GPU_IDS ${INDEX} GPU)
+        list(GET GPU_API_VERSIONS ${INDEX} API_VERSION)
+
+        # Extract API version
+        if(${API_VERSION} MATCHES "apiVersion[ ]*=[ ]*[0-9a-fA-F]*[ ]*[(]*([0-9]+[.][0-9]+[.][0-9]+)[)]*")
+            set(VULKAN_DRIVER_VERSION ${CMAKE_MATCH_1})
+        else()
+            message(FATAL_ERROR "API version match failed. This should not have happened...")
+        endif()
+
+        message(STATUS "${GPU} supports Vulkan API version '${VULKAN_DRIVER_VERSION}'.")
+
+        # Compare driver and header version
+        if(${VULKAN_DRIVER_VERSION} VERSION_LESS ${VULKAN_HEADER_VERSION})
+        # Version missmatch. Let us check if the minor version is the same.
+            if(${VULKAN_DRIVER_VERSION} MATCHES "[0-9]+[.]([0-9]+)[.][0-9]+")
+                set(VULKAN_DRIVER_MINOR_VERSION ${CMAKE_MATCH_1})
+            else()
+                message(FATAL_ERROR "Invalid Vulkan driver version '${VULKAN_DRIVER_VERSION}' found. Expected version in the following format: '[0-9]+.[0-9]+.[0-9]+'")
+            endif()
+            if(${VULKAN_HEADER_VERSION} MATCHES "[0-9]+[.]([0-9]+)[.][0-9]+")
+                set(VULKAN_HEADER_MINOR_VERSION ${CMAKE_MATCH_1})
+            else()
+                message(FATAL_ERROR "Invalid Vulkan Header version '${VULKAN_HEADER_VERSION}' found. Expected version in the following format: '[0-9]+.[0-9]+.[0-9]+'")
+            endif()
+
+            if(${VULKAN_DRIVER_MINOR_VERSION} EQUAL ${VULKAN_HEADER_MINOR_VERSION})
+                message(WARNING "Your GPU driver does not support Vulkan > ${VULKAN_DRIVER_VERSION}, but you try to use Vulkan Header ${VULKAN_HEADER_VERSION}. At least your driver supports the same minor version (${VULKAN_DRIVER_MINOR_VERSION}), so this should be fine but keep it in mind in case you encounter any strange behavior.")
+                set(VALID_GPU ${GPU})
+                set(VALID_VULKAN_VERSION ${VULKAN_DRIVER_VERSION})
+                break()
+            else()
+                message(STATUS "${GPU} does not support Vulkan > ${VULKAN_DRIVER_VERSION}.")
+            endif()
+        else()
+            set(VALID_GPU ${GPU})
+            set(VALID_VULKAN_VERSION ${VULKAN_DRIVER_VERSION})
+            break()
+        endif()
+    endforeach()
+
+    if("${VALID_GPU}" STREQUAL "")
+        message(FATAL_ERROR "None of your GPUs supports Vulkan Header ${VULKAN_HEADER_VERSION}. Please try updating your driver, or downgrade your Vulkan headers. If you know what you are doing, you can disable the Vulkan version check by setting 'KOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK' to 'ON' (-DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON).")
+    else()
+        message("Valid GPU (${VALID_GPU}) for Vulkan header version ${VULKAN_HEADER_VERSION} found. ${VALID_GPU} supports up to Vulkan ${VALID_VULKAN_VERSION}.")
+    endif()
+
+endfunction()

kompute/cmake/code_coverage.cmake

@@ -0,0 +1,35 @@
+# Code coverage
+set(CMAKE_BUILD_TYPE COVERAGE CACHE INTERNAL "Coverage build enabled")
+message(STATUS "Enabling gcov support")
+
+if(NOT "${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
+    set(COVERAGE_FLAG "--coverage")
+endif()
+
+set(CMAKE_CXX_FLAGS_COVERAGE
+    "-g -O0 ${COVERAGE_FLAG} -fprofile-arcs -ftest-coverage"
+    CACHE STRING "Flags used by the C++ compiler during coverage builds."
+    FORCE)
+set(CMAKE_C_FLAGS_COVERAGE
+    "-g -O0 ${COVERAGE_FLAG} -fprofile-arcs -ftest-coverage"
+    CACHE STRING "Flags used by the C compiler during coverage builds."
+    FORCE)
+set(CMAKE_EXE_LINKER_FLAGS_COVERAGE
+    ""
+    CACHE STRING "Flags used for linking binaries during coverage builds."
+    FORCE)
+set(CMAKE_SHARED_LINKER_FLAGS_COVERAGE
+    ""
+    CACHE STRING "Flags used by the shared libraries linker during coverage builds."
+    FORCE)
+
+set(CODECOV_DIR ${CMAKE_CURRENT_BINARY_DIR}/codecov/)
+set(CODECOV_DIR_LCOV ${CODECOV_DIR}lcov/)
+set(CODECOV_FILENAME_LCOV_INFO lcov.info)
+set(CODECOV_FILENAME_LCOV_INFO_FULL lcov_full.info)
+set(CODECOV_DIR_HTML ${CODECOV_DIR}html/)
+
+mark_as_advanced(CMAKE_CXX_FLAGS_COVERAGE
+    CMAKE_C_FLAGS_COVERAGE
+    CMAKE_EXE_LINKER_FLAGS_COVERAGE
+    CMAKE_SHARED_LINKER_FLAGS_COVERAGE)

kompute/cmake/deprecation_warnings.cmake

@@ -0,0 +1,15 @@
+if(KOMPUTE_OPT_REPO_SUBMODULE_BUILD)
+    message(FATAL_ERROR "'KOMPUTE_OPT_REPO_SUBMODULE_BUILD' got replaced by 'KOMPUTE_OPT_USE_BUILT_IN_SPDLOG', 'KOMPUTE_OPT_USE_BUILT_IN_FMT', 'KOMPUTE_OPT_USE_BUILT_IN_GOOGLE_TEST', 'KOMPUTE_OPT_USE_BUILT_IN_PYBIND11' and 'KOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER'. Please use them instead.")
+endif()
+
+if(KOMPUTE_OPT_BUILD_AS_SHARED_LIB)
+    message(FATAL_ERROR "'KOMPUTE_OPT_BUILD_AS_SHARED_LIB' is deprecated and should not be used. Instead use the default 'BUILD_SHARED_LIBS' CMake switch.")
+endif()
+
+if(KOMPUTE_OPT_BUILD_SINGLE_HEADER)
+    message(FATAL_ERROR "'KOMPUTE_OPT_BUILD_SINGLE_HEADER' is deprecated and should not be used. The single header will now always be build and can be included via '#include<kompute/kompute.h>'.")
+endif()
+
+if(KOMPUTE_OPT_ENABLE_SPDLOG)
+    message(FATAL_ERROR "'KOMPUTE_OPT_ENABLE_SPDLOG' is deprecated and should not be used. It got replaced by 'KOMPUTE_OPT_LOG_LEVEL'. This option can be set to a variety of log levels (e.g. 'Off', 'Trace', 'Debug', 'Default', ...).")
+endif()

kompute/cmake/komputeConfig.cmake.in

@@ -0,0 +1,8 @@
+include(CMakeFindDependencyMacro)
+@PACKAGE_INIT@
+
+find_dependency(VULKAN REQUIRED)
+
+include(${CMAKE_CURRENT_LIST_DIR}/komputeTargets.cmake)
+
+check_required_components(kompute)

kompute/cmake/vulkan_shader_compiler.cmake

@@ -0,0 +1,43 @@
+function(vulkan_compile_shader)
+     find_program(GLS_LANG_VALIDATOR_PATH NAMES glslangValidator)
+     if(GLS_LANG_VALIDATOR_PATH STREQUAL "GLS_LANG_VALIDATOR_PATH-NOTFOUND")
+          message(FATAL_ERROR "glslangValidator not found.")
+          return()
+     endif()
+
+     cmake_parse_arguments(SHADER_COMPILE "" "INFILE;OUTFILE;NAMESPACE;RELATIVE_PATH" "" ${ARGN})
+     set(SHADER_COMPILE_INFILE_FULL "${CMAKE_CURRENT_SOURCE_DIR}/${SHADER_COMPILE_INFILE}")
+     set(SHADER_COMPILE_SPV_FILE_FULL "${CMAKE_CURRENT_BINARY_DIR}/${SHADER_COMPILE_INFILE}.spv")
+     set(SHADER_COMPILE_HEADER_FILE_FULL "${CMAKE_CURRENT_BINARY_DIR}/${SHADER_COMPILE_OUTFILE}")
+
+     if(NOT SHADER_COMPILE_RELATIVE_PATH)
+          set(SHADER_COMPILE_RELATIVE_PATH "${PROJECT_SOURCE_DIR}/cmake")
+     endif()
+    
+     # .comp -> .spv
+     add_custom_command(OUTPUT "${SHADER_COMPILE_SPV_FILE_FULL}"
+                        COMMAND "${GLS_LANG_VALIDATOR_PATH}"
+                        ARGS "-V"
+                             "${SHADER_COMPILE_INFILE_FULL}"
+                             "-o"
+                             "${SHADER_COMPILE_SPV_FILE_FULL}"
+                        COMMENT "Compile vulkan compute shader from file '${SHADER_COMPILE_INFILE_FULL}' to '${SHADER_COMPILE_SPV_FILE_FULL}'."
+                        MAIN_DEPENDENCY "${SHADER_COMPILE_INFILE_FULL}")
+
+     # Check if big or little endian
+     include (TestBigEndian)
+     TEST_BIG_ENDIAN(IS_BIG_ENDIAN)
+
+     # .spv -> .hpp
+     add_custom_command(OUTPUT "${SHADER_COMPILE_HEADER_FILE_FULL}"
+                        COMMAND ${CMAKE_COMMAND}
+                        ARGS "-DINPUT_SHADER_FILE=${SHADER_COMPILE_SPV_FILE_FULL}"
+                             "-DOUTPUT_HEADER_FILE=${SHADER_COMPILE_HEADER_FILE_FULL}"
+                             "-DHEADER_NAMESPACE=${SHADER_COMPILE_NAMESPACE}"
+                             "-DIS_BIG_ENDIAN=${IS_BIG_ENDIAN}"
+                             "-P"
+                             "${SHADER_COMPILE_RELATIVE_PATH}/bin_file_to_header.cmake"
+                        WORKING_DIRECTORY "${SHADER_COMPILE_RELATIVE_PATH}"
+                        COMMENT "Converting compiled shader '${SHADER_COMPILE_SPV_FILE_FULL}' to header file '${SHADER_COMPILE_HEADER_FILE_FULL}'."
+                        MAIN_DEPENDENCY "${SHADER_COMPILE_SPV_FILE_FULL}")
+endfunction()

kompute/config/FindSphinx.cmake

@@ -0,0 +1,16 @@
+# Look for an executable called sphinx-build
+find_program(SPHINX_EXECUTABLE
+    NAMES sphinx-build
+    DOC "Path to sphinx-build executable")
+
+if(SPHINX_EXECUTABLE STREQUAL "SPHINX_EXECUTABLE-NOTFOUND")
+    message(FATAL_ERROR "sphinx-build not found.")
+endif()
+
+include(FindPackageHandleStandardArgs)
+
+# Handle standard arguments to find_package like REQUIRED and QUIET
+find_package_handle_standard_args(
+    Sphinx
+    "Failed to find sphinx-build executable"
+    SPHINX_EXECUTABLE)

kompute/external/bin/xxd.c

@@ -0,0 +1,819 @@
+/*
+As indicated at https://lists.debian.org/debian-legal/2015/01/msg00037.html,
+the author has permitted redistribution of xxd under the MIT license, as follows:
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be included
+in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * xxd: my hexdump facility. jw
+ *
+ *  2.10.90 changed to word output
+ *  3.03.93 new indent style, dumb bug inserted and fixed.
+ *	    -c option, mls
+ * 26.04.94 better option parser, -ps, -l, -s added.
+ *  1.07.94 -r badly needs - as input file.  Per default autoskip over
+ *	       consecutive lines of zeroes, as unix od does.
+ *	    -a shows them too.
+ *	    -i dump as c-style #include "file.h"
+ *  1.11.95 if "xxd -i" knows the filename, an 'unsigned char filename_bits[]'
+ *	    array is written in correct c-syntax.
+ *	    -s improved, now defaults to absolute seek, relative requires a '+'.
+ *	    -r improved, now -r -s -0x... is supported.
+ *	       change/suppress leading '\0' bytes.
+ *	    -l n improved: stops exactly after n bytes.
+ *	    -r improved, better handling of partial lines with trailing garbage.
+ *	    -r improved, now -r -p works again!
+ *	    -r improved, less flushing, much faster now! (that was silly)
+ *  3.04.96 Per repeated request of a single person: autoskip defaults to off.
+ * 15.05.96 -v added. They want to know the version.
+ *	    -a fixed, to show last line inf file ends in all zeros.
+ *	    -u added: Print upper case hex-letters, as preferred by unix bc.
+ *	    -h added to usage message. Usage message extended.
+ *	    Now using outfile if specified even in normal mode, aehem.
+ *	    No longer mixing of ints and longs. May help doze people.
+ *	    Added binify ioctl for same reason. (Enough Doze stress for 1996!)
+ * 16.05.96 -p improved, removed occasional superfluous linefeed.
+ * 20.05.96 -l 0 fixed. tried to read anyway.
+ * 21.05.96 -i fixed. now honours -u, and prepends __ to numeric filenames.
+ *	    compile -DWIN32 for NT or W95. George V. Reilly, * -v improved :-)
+ *	    support --gnuish-longhorn-options
+ * 25.05.96 MAC support added: CodeWarrior already uses ``outline'' in Types.h
+ *	    which is included by MacHeaders (Axel Kielhorn). Renamed to
+ *	    xxdline().
+ *  7.06.96 -i printed 'int' instead of 'char'. *blush*
+ *	    added Bram's OS2 ifdefs...
+ * 18.07.96 gcc -Wall @ SunOS4 is now slient.
+ *	    Added osver for MSDOS/DJGPP/WIN32.
+ * 29.08.96 Added size_t to strncmp() for Amiga.
+ * 24.03.97 Windows NT support (Phil Hanna). Clean exit for Amiga WB (Bram)
+ * 02.04.97 Added -E option, to have EBCDIC translation instead of ASCII
+ *	    (azc10@yahoo.com)
+ * 22.05.97 added -g (group octets) option (jcook@namerica.kla.com).
+ * 23.09.98 nasty -p -r misfeature fixed: slightly wrong output, when -c was
+ *	    missing or wrong.
+ * 26.09.98 Fixed: 'xxd -i infile outfile' did not truncate outfile.
+ * 27.10.98 Fixed: -g option parser required blank.
+ *	    option -b added: 01000101 binary output in normal format.
+ * 16.05.00 Added VAXC changes by Stephen P. Wall
+ * 16.05.00 Improved MMS file and merge for VMS by Zoltan Arpadffy
+ *
+ * (c) 1990-1998 by Juergen Weigert (jnweiger@informatik.uni-erlangen.de)
+ *
+ * Small changes made afterwards by Bram Moolenaar et al.
+ *
+ * Distribute freely and credit me,
+ * make money and share with me,
+ * lose money and don't ask me.
+ *
+ *
+ */
+
+/* Visual Studio 2005 has 'deprecated' many of the standard CRT functions */
+#if _MSC_VER >= 1400
+# define _CRT_SECURE_NO_DEPRECATE
+# define _CRT_NONSTDC_NO_DEPRECATE
+#endif
+
+#include <stdio.h>
+#ifdef VAXC
+# include <file.h>
+#else
+# include <fcntl.h>
+#endif
+#ifdef __TSC__
+# define MSDOS
+#endif
+#if !defined(OS2) && defined(__EMX__)
+# define OS2
+#endif
+#if defined(MSDOS) || defined(WIN32) || defined(OS2) || defined(__BORLANDC__) || defined(CYGWIN)
+# include <io.h>	/* for setmode() */
+#else
+# ifdef UNIX
+#  include <unistd.h>
+# endif
+#endif
+#include <stdlib.h>
+#include <string.h>	/* for strncmp() */
+#include <ctype.h>	/* for isalnum() */
+#if __MWERKS__ && !defined(BEBOX)
+# include <unix.h>	/* for fdopen() on MAC */
+#endif
+
+#if defined(__BORLANDC__) && __BORLANDC__ <= 0x0410 && !defined(fileno)
+/* Missing define and prototype grabbed from the BC 4.0 <stdio.h> */
+# define fileno(f)       ((f)->fd)
+FILE   _FAR *_Cdecl _FARFUNC fdopen(int __handle, char _FAR *__type);
+#endif
+
+
+/*  This corrects the problem of missing prototypes for certain functions
+ *  in some GNU installations (e.g. SunOS 4.1.x).
+ *  Darren Hiebert <darren@hmi.com> (sparc-sun-sunos4.1.3_U1/2.7.2.2)
+ */
+#if defined(__GNUC__) && defined(__STDC__)
+# ifndef __USE_FIXED_PROTOTYPES__
+#  define __USE_FIXED_PROTOTYPES__
+# endif
+#endif
+
+#ifndef __USE_FIXED_PROTOTYPES__
+/*
+ * This is historic and works only if the compiler really has no prototypes:
+ *
+ * Include prototypes for Sun OS 4.x, when using an ANSI compiler.
+ * FILE is defined on OS 4.x, not on 5.x (Solaris).
+ * if __SVR4 is defined (some Solaris versions), don't include this.
+ */
+#if defined(sun) && defined(FILE) && !defined(__SVR4) && defined(__STDC__)
+#  define __P(a) a
+/* excerpt from my sun_stdlib.h */
+extern int fprintf __P((FILE *, char *, ...));
+extern int fputs   __P((char *, FILE *));
+extern int _flsbuf __P((unsigned char, FILE *));
+extern int _filbuf __P((FILE *));
+extern int fflush  __P((FILE *));
+extern int fclose  __P((FILE *));
+extern int fseek   __P((FILE *, long, int));
+extern int rewind  __P((FILE *));
+
+extern void perror __P((char *));
+# endif
+#endif
+
+extern long int strtol();
+extern long int ftell();
+
+char version[] = "xxd V1.10 27oct98 by Juergen Weigert";
+#ifdef WIN32
+char osver[] = " (Win32)";
+#else
+# ifdef DJGPP
+char osver[] = " (dos 32 bit)";
+# else
+#  ifdef MSDOS
+char osver[] = " (dos 16 bit)";
+#  else
+char osver[] = "";
+#  endif
+# endif
+#endif
+
+#if !defined(CYGWIN) && (defined(CYGWIN32) || defined(__CYGWIN__) || defined(__CYGWIN32__))
+# define CYGWIN
+#endif
+#if defined(MSDOS) || defined(WIN32) || defined(OS2)
+# define BIN_READ(yes)  ((yes) ? "rb" : "rt")
+# define BIN_WRITE(yes) ((yes) ? "wb" : "wt")
+# define BIN_CREAT(yes) ((yes) ? (O_CREAT|O_BINARY) : O_CREAT)
+# define BIN_ASSIGN(fp, yes) setmode(fileno(fp), (yes) ? O_BINARY : O_TEXT)
+# define PATH_SEP '\\'
+#elif defined(CYGWIN)
+# define BIN_READ(yes)  ((yes) ? "rb" : "rt")
+# define BIN_WRITE(yes) ((yes) ? "wb" : "w")
+# define BIN_CREAT(yes) ((yes) ? (O_CREAT|O_BINARY) : O_CREAT)
+# define BIN_ASSIGN(fp, yes) ((yes) ? (void) setmode(fileno(fp), O_BINARY) : (void) (fp))
+# define PATH_SEP '/'
+#else
+# ifdef VMS
+#  define BIN_READ(dummy)  "r"
+#  define BIN_WRITE(dummy) "w"
+#  define BIN_CREAT(dummy) O_CREAT
+#  define BIN_ASSIGN(fp, dummy) fp
+#  define PATH_SEP ']'
+#  define FILE_SEP '.'
+# else
+#  define BIN_READ(dummy)  "r"
+#  define BIN_WRITE(dummy) "w"
+#  define BIN_CREAT(dummy) O_CREAT
+#  define BIN_ASSIGN(fp, dummy) fp
+#  define PATH_SEP '/'
+# endif
+#endif
+
+/* open has only to arguments on the Mac */
+#if __MWERKS__
+# define OPEN(name, mode, umask) open(name, mode)
+#else
+# define OPEN(name, mode, umask) open(name, mode, umask)
+#endif
+
+#ifdef AMIGA
+# define STRNCMP(s1, s2, l) strncmp(s1, s2, (size_t)l)
+#else
+# define STRNCMP(s1, s2, l) strncmp(s1, s2, l)
+#endif
+
+#ifndef __P
+# if defined(__STDC__) || defined(MSDOS) || defined(WIN32) || defined(OS2) \
+        || defined(__BORLANDC__)
+#  define __P(a) a
+# else
+#  define __P(a) ()
+# endif
+#endif
+
+/* Let's collect some prototypes */
+/* CodeWarrior is really picky about missing prototypes */
+static void exit_with_usage __P((char *));
+static int huntype __P((FILE *, FILE *, FILE *, char *, int, int, long));
+static void xxdline __P((FILE *, char *, int));
+
+#define TRY_SEEK	/* attempt to use lseek, or skip forward by reading */
+#define COLS 256	/* change here, if you ever need more columns */
+#define LLEN (11 + (9*COLS-1)/1 + COLS + 2)
+
+char hexxa[] = "0123456789abcdef0123456789ABCDEF", *hexx = hexxa;
+
+/* the different hextypes known by this program: */
+#define HEX_NORMAL 0
+#define HEX_POSTSCRIPT 1
+#define HEX_CINCLUDE 2
+#define HEX_BITS 3		/* not hex a dump, but bits: 01111001 */
+
+static void
+exit_with_usage(pname)
+char *pname;
+{
+  fprintf(stderr, "Usage:\n       %s [options] [infile [outfile]]\n", pname);
+  fprintf(stderr, "    or\n       %s -r [-s [-]offset] [-c cols] [-ps] [infile [outfile]]\n", pname);
+  fprintf(stderr, "Options:\n");
+  fprintf(stderr, "    -a          toggle autoskip: A single '*' replaces nul-lines. Default off.\n");
+  fprintf(stderr, "    -b          binary digit dump (incompatible with -p,-i,-r). Default hex.\n");
+  fprintf(stderr, "    -c cols     format <cols> octets per line. Default 16 (-i: 12, -ps: 30).\n");
+  fprintf(stderr, "    -E          show characters in EBCDIC. Default ASCII.\n");
+  fprintf(stderr, "    -g          number of octets per group in normal output. Default 2.\n");
+  fprintf(stderr, "    -h          print this summary.\n");
+  fprintf(stderr, "    -i          output in C include file style.\n");
+  fprintf(stderr, "    -l len      stop after <len> octets.\n");
+  fprintf(stderr, "    -ps         output in postscript plain hexdump style.\n");
+  fprintf(stderr, "    -r          reverse operation: convert (or patch) hexdump into binary.\n");
+  fprintf(stderr, "    -r -s off   revert with <off> added to file positions found in hexdump.\n");
+  fprintf(stderr, "    -s %sseek  start at <seek> bytes abs. %sinfile offset.\n",
+#ifdef TRY_SEEK
+      "[+][-]", "(or +: rel.) ");
+#else
+      "", "");
+#endif
+  fprintf(stderr, "    -u          use upper case hex letters.\n");
+  fprintf(stderr, "    -v          show version: \"%s%s\".\n", version, osver);
+  exit(1);
+}
+
+/*
+ * Max. cols binary characters are decoded from the input stream per line.
+ * Two adjacent garbage characters after evaluated data delimit valid data.
+ * Everything up to the next newline is discarded.
+ *
+ * The name is historic and came from 'undo type opt h'.
+ */
+static int
+huntype(fpi, fpo, fperr, pname, cols, hextype, base_off)
+FILE *fpi, *fpo, *fperr;
+char *pname;
+int cols, hextype;
+long base_off;
+{
+  int c, ign_garb = 1, n1 = -1, n2 = 0, n3, p = cols;
+  long have_off = 0, want_off = 0;
+
+  rewind(fpi);
+
+  while ((c = getc(fpi)) != EOF)
+    {
+      if (c == '\r')	/* Doze style input file? */
+    continue;
+
+#if 0	/* this doesn't work when there is normal text after the hex codes in
+       the last line that looks like hex */
+      if (c == ' ' || c == '\n' || c == '\t')  /* allow multiple spaces */
+    continue;
+#endif
+
+      n3 = n2;
+      n2 = n1;
+
+      if (c >= '0' && c <= '9')
+    n1 = c - '0';
+      else if (c >= 'a' && c <= 'f')
+    n1 = c - 'a' + 10;
+      else if (c >= 'A' && c <= 'F')
+    n1 = c - 'A' + 10;
+      else
+    {
+      n1 = -1;
+      if (ign_garb)
+        continue;
+    }
+
+      ign_garb = 0;
+
+      if (p >= cols)
+    {
+      if (!hextype)
+        {
+          if (n1 < 0)
+        {
+          p = 0;
+          continue;
+        }
+          want_off = (want_off << 4) | n1;
+          continue;
+        }
+      else
+        p = 0;
+    }
+
+      if (base_off + want_off != have_off)
+    {
+      fflush(fpo);
+#ifdef TRY_SEEK
+      c = fseek(fpo, base_off + want_off - have_off, 1);
+      if (c >= 0)
+        have_off = base_off + want_off;
+#endif
+      if (base_off + want_off < have_off)
+        {
+          fprintf(fperr, "%s: sorry, cannot seek backwards.\n", pname);
+          return 5;
+        }
+      for (; have_off < base_off + want_off; have_off++)
+        putc(0, fpo);
+    }
+
+      if (n2 >= 0 && n1 >= 0)
+    {
+      putc((n2 << 4) | n1, fpo);
+      have_off++;
+      want_off++;
+      n1 = -1;
+      if ((++p >= cols) && !hextype)
+        {
+          /* skip rest of line as garbage */
+          want_off = 0;
+          while ((c = getc(fpi)) != '\n' && c != EOF)
+        ;
+          ign_garb = 1;
+        }
+    }
+      else if (n1 < 0 && n2 < 0 && n3 < 0)
+    {
+      /* already stumbled into garbage, skip line, wait and see */
+      if (!hextype)
+        want_off = 0;
+      while ((c = getc(fpi)) != '\n' && c != EOF)
+        ;
+      ign_garb = 1;
+    }
+    }
+  fflush(fpo);
+#ifdef TRY_SEEK
+  fseek(fpo, 0L, 2);
+#endif
+  fclose(fpo);
+  fclose(fpi);
+  return 0;
+}
+
+/*
+ * Print line l. If nz is false, xxdline regards the line a line of
+ * zeroes. If there are three or more consecutive lines of zeroes,
+ * they are replaced by a single '*' character.
+ *
+ * If the output ends with more than two lines of zeroes, you
+ * should call xxdline again with l being the last line and nz
+ * negative. This ensures that the last line is shown even when
+ * it is all zeroes.
+ *
+ * If nz is always positive, lines are never suppressed.
+ */
+static void
+xxdline(fp, l, nz)
+FILE *fp;
+char *l;
+int nz;
+{
+  static char z[LLEN+1];
+  static int zero_seen = 0;
+
+  if (!nz && zero_seen == 1)
+    strcpy(z, l);
+
+  if (nz || !zero_seen++)
+    {
+      if (nz)
+    {
+      if (nz < 0)
+        zero_seen--;
+      if (zero_seen == 2)
+        fputs(z, fp);
+      if (zero_seen > 2)
+        fputs("*\n", fp);
+    }
+      if (nz >= 0 || zero_seen > 0)
+    fputs(l, fp);
+      if (nz)
+    zero_seen = 0;
+    }
+}
+
+/* This is an EBCDIC to ASCII conversion table */
+/* from a proposed BTL standard April 16, 1979 */
+static unsigned char etoa64[] =
+{
+    0040,0240,0241,0242,0243,0244,0245,0246,
+    0247,0250,0325,0056,0074,0050,0053,0174,
+    0046,0251,0252,0253,0254,0255,0256,0257,
+    0260,0261,0041,0044,0052,0051,0073,0176,
+    0055,0057,0262,0263,0264,0265,0266,0267,
+    0270,0271,0313,0054,0045,0137,0076,0077,
+    0272,0273,0274,0275,0276,0277,0300,0301,
+    0302,0140,0072,0043,0100,0047,0075,0042,
+    0303,0141,0142,0143,0144,0145,0146,0147,
+    0150,0151,0304,0305,0306,0307,0310,0311,
+    0312,0152,0153,0154,0155,0156,0157,0160,
+    0161,0162,0136,0314,0315,0316,0317,0320,
+    0321,0345,0163,0164,0165,0166,0167,0170,
+    0171,0172,0322,0323,0324,0133,0326,0327,
+    0330,0331,0332,0333,0334,0335,0336,0337,
+    0340,0341,0342,0343,0344,0135,0346,0347,
+    0173,0101,0102,0103,0104,0105,0106,0107,
+    0110,0111,0350,0351,0352,0353,0354,0355,
+    0175,0112,0113,0114,0115,0116,0117,0120,
+    0121,0122,0356,0357,0360,0361,0362,0363,
+    0134,0237,0123,0124,0125,0126,0127,0130,
+    0131,0132,0364,0365,0366,0367,0370,0371,
+    0060,0061,0062,0063,0064,0065,0066,0067,
+    0070,0071,0372,0373,0374,0375,0376,0377
+};
+
+const char* extract_filename(const char* path) {
+    const char* filename = strrchr(path, '/');
+    if (filename) {
+        return filename + 1;
+    }
+    return path;
+}
+
+int
+main(argc, argv)
+int argc;
+char *argv[];
+{
+  FILE *fp, *fpo;
+  int c, e, p = 0, relseek = 1, negseek = 0, revert = 0;
+  int cols = 0, nonzero = 0, autoskip = 0, hextype = HEX_NORMAL;
+  int ebcdic = 0;
+  int octspergrp = -1;	/* number of octets grouped in output */
+  int grplen;		/* total chars per octet group */
+  long length = -1, n = 0, seekoff = 0;
+  char l[LLEN+1];
+  char *pname, *pp;
+
+#ifdef AMIGA
+  /* This program doesn't work when started from the Workbench */
+  if (argc == 0)
+    exit(1);
+#endif
+
+  pname = argv[0];
+  for (pp = pname; *pp; )
+    if (*pp++ == PATH_SEP)
+      pname = pp;
+#ifdef FILE_SEP
+  for (pp = pname; *pp; pp++)
+    if (*pp == FILE_SEP)
+      {
+    *pp = '\0';
+    break;
+      }
+#endif
+
+  while (argc >= 2)
+    {
+      pp = argv[1] + (!STRNCMP(argv[1], "--", 2) && argv[1][2]);
+       if (!STRNCMP(pp, "-a", 2)) autoskip = 1 - autoskip;
+      else if (!STRNCMP(pp, "-b", 2)) hextype = HEX_BITS;
+      else if (!STRNCMP(pp, "-u", 2)) hexx = hexxa + 16;
+      else if (!STRNCMP(pp, "-p", 2)) hextype = HEX_POSTSCRIPT;
+      else if (!STRNCMP(pp, "-i", 2)) hextype = HEX_CINCLUDE;
+      else if (!STRNCMP(pp, "-r", 2)) revert++;
+      else if (!STRNCMP(pp, "-E", 2)) ebcdic++;
+      else if (!STRNCMP(pp, "-v", 2))
+    {
+      fprintf(stderr, "%s%s\n", version, osver);
+      exit(0);
+    }
+      else if (!STRNCMP(pp, "-c", 2))
+    {
+      if (pp[2] && STRNCMP("ols", pp + 2, 3))
+        cols = (int)strtol(pp + 2, NULL, 0);
+      else
+        {
+          if (!argv[2])
+        exit_with_usage(pname);
+          cols = (int)strtol(argv[2], NULL, 0);
+          argv++;
+          argc--;
+        }
+    }
+      else if (!STRNCMP(pp, "-g", 2))
+    {
+      if (pp[2] && STRNCMP("roupsize", pp + 2, 8))
+        octspergrp = (int)strtol(pp + 2, NULL, 0);
+      else
+        {
+          if (!argv[2])
+        exit_with_usage(pname);
+          octspergrp = (int)strtol(argv[2], NULL, 0);
+          argv++;
+          argc--;
+        }
+    }
+      else if (!STRNCMP(pp, "-s", 2))
+    {
+      relseek = 0;
+      negseek = 0;
+      if (pp[2] && STRNCMP("kip", pp+2, 3) && STRNCMP("eek", pp+2, 3))
+        {
+#ifdef TRY_SEEK
+          if (pp[2] == '+')
+        relseek++;
+          if (pp[2+relseek] == '-')
+        negseek++;
+#endif
+          seekoff = strtol(pp + 2+relseek+negseek, (char **)NULL, 0);
+        }
+      else
+        {
+          if (!argv[2])
+        exit_with_usage(pname);
+#ifdef TRY_SEEK
+          if (argv[2][0] == '+')
+        relseek++;
+          if (argv[2][relseek] == '-')
+        negseek++;
+#endif
+          seekoff = strtol(argv[2] + relseek+negseek, (char **)NULL, 0);
+          argv++;
+          argc--;
+        }
+    }
+      else if (!STRNCMP(pp, "-l", 2))
+    {
+      if (pp[2] && STRNCMP("en", pp + 2, 2))
+        length = strtol(pp + 2, (char **)NULL, 0);
+      else
+        {
+          if (!argv[2])
+        exit_with_usage(pname);
+          length = strtol(argv[2], (char **)NULL, 0);
+          argv++;
+          argc--;
+        }
+    }
+      else if (!strcmp(pp, "--"))	/* end of options */
+    {
+      argv++;
+      argc--;
+      break;
+    }
+      else if (pp[0] == '-' && pp[1])	/* unknown option */
+    exit_with_usage(pname);
+      else
+    break;				/* not an option */
+
+      argv++;				/* advance to next argument */
+      argc--;
+    }
+
+  if (!cols)
+    switch (hextype)
+      {
+      case HEX_POSTSCRIPT:	cols = 30; break;
+      case HEX_CINCLUDE:	cols = 12; break;
+      case HEX_BITS:		cols = 6; break;
+      case HEX_NORMAL:
+      default:			cols = 16; break;
+      }
+
+  if (octspergrp < 0)
+    switch (hextype)
+      {
+      case HEX_BITS:		octspergrp = 1; break;
+      case HEX_NORMAL:		octspergrp = 2; break;
+      case HEX_POSTSCRIPT:
+      case HEX_CINCLUDE:
+      default:			octspergrp = 0; break;
+      }
+
+  if (cols < 1 || ((hextype == HEX_NORMAL || hextype == HEX_BITS)
+                                && (cols > COLS)))
+    {
+      fprintf(stderr, "%s: invalid number of columns (max. %d).\n", pname, COLS);
+      exit(1);
+    }
+
+  if (octspergrp < 1)
+    octspergrp = cols;
+
+  if (argc > 3)
+    exit_with_usage(pname);
+
+  if (argc == 1 || (argv[1][0] == '-' && !argv[1][1]))
+    BIN_ASSIGN(fp = stdin, !revert);
+  else
+    {
+      if ((fp = fopen(argv[1], BIN_READ(!revert))) == NULL)
+    {
+      fprintf(stderr,"%s: ", pname);
+      perror(argv[1]);
+      return 2;
+    }
+    }
+
+  if (argc < 3 || (argv[2][0] == '-' && !argv[2][1]))
+    BIN_ASSIGN(fpo = stdout, revert);
+  else
+    {
+      int fd;
+      int mode = revert ? O_WRONLY : (O_TRUNC|O_WRONLY);
+
+      if (((fd = OPEN(argv[2], mode | BIN_CREAT(revert), 0666)) < 0) ||
+      (fpo = fdopen(fd, BIN_WRITE(revert))) == NULL)
+    {
+      fprintf(stderr, "%s: ", pname);
+      perror(argv[2]);
+      return 3;
+    }
+      rewind(fpo);
+    }
+
+  if (revert)
+    {
+      if (hextype && (hextype != HEX_POSTSCRIPT))
+    {
+      fprintf(stderr, "%s: sorry, cannot revert this type of hexdump\n", pname);
+      return -1;
+    }
+      return huntype(fp, fpo, stderr, pname, cols, hextype,
+        negseek ? -seekoff : seekoff);
+    }
+
+  if (seekoff || negseek || !relseek)
+    {
+#ifdef TRY_SEEK
+      if (relseek)
+    e = fseek(fp, negseek ? -seekoff : seekoff, 1);
+      else
+    e = fseek(fp, negseek ? -seekoff : seekoff, negseek ? 2 : 0);
+      if (e < 0 && negseek)
+    {
+      fprintf(stderr, "%s: sorry cannot seek.\n", pname);
+      return 4;
+    }
+      if (e >= 0)
+    seekoff = ftell(fp);
+      else
+#endif
+    {
+      long s = seekoff;
+
+      while (s--)
+        (void)getc(fp);
+    }
+    }
+
+  if (hextype == HEX_CINCLUDE)
+    {
+      const char* filename = extract_filename(argv[1]);
+
+      if (fp != stdin)
+    {
+      fprintf(fpo, "unsigned char %s", isdigit((int)filename[0]) ? "__" : "");
+      for (e = 0; (c = filename[e]) != 0; e++)
+        putc(isalnum(c) ? c : '_', fpo);
+      fputs("[] = {\n", fpo);
+    }
+
+      p = 0;
+      while ((length < 0 || p < length) && (c = getc(fp)) != EOF)
+    {
+      fprintf(fpo, (hexx == hexxa) ? "%s0x%02x" : "%s0X%02X",
+        (p % cols) ? ", " : ",\n  "+2*!p,  c);
+      p++;
+    }
+
+      if (p)
+    fputs("\n};\n"+3*(fp == stdin), fpo);
+
+      if (fp != stdin)
+    {
+      fprintf(fpo, "unsigned int %s", isdigit((int)filename[0]) ? "__" : "");
+      for (e = 0; (c = filename[e]) != 0; e++)
+        putc(isalnum(c) ? c : '_', fpo);
+      fprintf(fpo, "_len = %d;\n", p);
+    }
+
+      fclose(fp);
+      fclose(fpo);
+      return 0;
+    }
+
+  if (hextype == HEX_POSTSCRIPT)
+    {
+      p = cols;
+      while ((length < 0 || n < length) && (e = getc(fp)) != EOF)
+    {
+      putchar(hexx[(e >> 4) & 0xf]);
+      putchar(hexx[(e     ) & 0xf]);
+      n++;
+      if (!--p)
+        {
+          putchar('\n');
+          p = cols;
+        }
+    }
+      if (p < cols)
+    putchar('\n');
+      fclose(fp);
+      fclose(fpo);
+      return 0;
+    }
+
+  /* hextype: HEX_NORMAL or HEX_BITS */
+
+  if (hextype == HEX_NORMAL)
+    grplen = octspergrp + octspergrp + 1;	/* chars per octet group */
+  else	/* hextype == HEX_BITS */
+    grplen = 8 * octspergrp + 1;
+
+  while ((length < 0 || n < length) && (e = getc(fp)) != EOF)
+    {
+      if (p == 0)
+    {
+      sprintf(l, "%07lx: ", n + seekoff);
+      for (c = 9; c < LLEN; l[c++] = ' ');
+    }
+      if (hextype == HEX_NORMAL)
+    {
+      l[c = (9 + (grplen * p) / octspergrp)] = hexx[(e >> 4) & 0xf];
+      l[++c]			       = hexx[ e       & 0xf];
+    }
+      else /* hextype == HEX_BITS */
+    {
+      int i;
+
+      c = (9 + (grplen * p) / octspergrp) - 1;
+      for (i = 7; i >= 0; i--)
+        l[++c] = (e & (1 << i)) ? '1' : '0';
+    }
+      if (ebcdic)
+    e = (e < 64) ? '.' : etoa64[e-64];
+      /* When changing this update definition of LLEN above. */
+      l[11 + (grplen * cols - 1)/octspergrp + p] =
+#ifdef __MVS__
+      (e >= 64)
+#else
+      (e > 31 && e < 127)
+#endif
+      ? e : '.';
+      if (e)
+    nonzero++;
+      n++;
+      if (++p == cols)
+    {
+      l[c = (11 + (grplen * cols - 1)/octspergrp + p)] = '\n'; l[++c] = '\0';
+      xxdline(fpo, l, autoskip ? nonzero : 1);
+      nonzero = 0;
+      p = 0;
+    }
+    }
+  if (p)
+    {
+      l[c = (11 + (grplen * cols - 1)/octspergrp + p)] = '\n'; l[++c] = '\0';
+      xxdline(fpo, l, 1);
+    }
+  else if (autoskip)
+    xxdline(fpo, l, -1);	/* last chance to flush out suppressed lines */
+
+  fclose(fp);
+  fclose(fpo);
+  return 0;
+}

kompute/kompute-config.cmake

@@ -0,0 +1,28 @@
+# General purpose GPU compute framework built on Vulkan to
+# support 1000s of cross vendor graphics cards
+# (AMD, Qualcomm, NVIDIA & friends). Blazing fast, mobile-enabled,
+# asynchronous and optimized for advanced GPU data processing use cases.
+# Backed by the Linux Foundation. 
+#
+# Finding this module will define the following variables:
+#  KOMPUTE_FOUND - True if the core library has been found
+#  KOMPUTE_LIBRARIES - Path to the core library archive
+#  KOMPUTE_INCLUDE_DIRS - Path to the include directories. Gives access
+#                     to kompute.h, as a single include which must be included in every
+#                     file that uses this interface. Else it also points to the
+#                     directory for individual includes.
+
+find_path(KOMPUTE_INCLUDE_DIR
+          NAMES kompute.h)
+
+find_library(KOMPUTE_LIBRARY
+             NAMES kompute
+             HINTS ${KOMPUTE_LIBRARY_ROOT})
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(KOMPUTE REQUIRED_VARS KOMPUTE_LIBRARY KOMPUTE_INCLUDE_DIR)
+
+if(KOMPUTE_FOUND)
+    set(KOMPUTE_LIBRARIES ${KOMPUTE_LIBRARY})
+    set(KOMPUTE_INCLUDE_DIRS ${KOMPUTE_INCLUDE_DIR})
+endif()

kompute/op_add.comp

@@ -0,0 +1,145 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    uint row;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+
+    out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i) + pcs.inBOff];
+}

kompute/op_addrow.comp

@@ -0,0 +1,145 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    uint row;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+
+    out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
+}

kompute/op_cpy_f16_f16.comp

@@ -0,0 +1,176 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+#define nth 32
+#define IN_TYPE float16_t
+#define IN_TYPE_SIZE 2
+#define OUT_TYPE float16_t
+#define OUT_TYPE_SIZE 2
+
+layout(local_size_x = nth) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute/op_cpy_f16_f32.comp

@@ -0,0 +1,176 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+#define nth 32
+#define IN_TYPE float16_t
+#define IN_TYPE_SIZE 2
+#define OUT_TYPE float
+#define OUT_TYPE_SIZE 4
+
+layout(local_size_x = nth) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute/op_cpy_f32_f16.comp

@@ -0,0 +1,176 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+#define nth 32
+#define IN_TYPE float
+#define IN_TYPE_SIZE 4
+#define OUT_TYPE float16_t
+#define OUT_TYPE_SIZE 2
+
+layout(local_size_x = nth) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute/op_cpy_f32_f32.comp

@@ -0,0 +1,168 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+#define nth 32
+#define IN_TYPE float
+#define IN_TYPE_SIZE 4
+#define OUT_TYPE float
+#define OUT_TYPE_SIZE 4
+
+layout(local_size_x = nth) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute/op_diagmask.comp

@@ -0,0 +1,153 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint n_past;
+    int ne00;
+    int ne01;
+} pcs;
+
+void main() {
+    const uint i02 = gl_WorkGroupID.z;
+    const uint i01 = gl_WorkGroupID.y;
+    const uint i00 = gl_WorkGroupID.x;
+
+    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
+
+    if (i00 > pcs.n_past + i01) {
+        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
+    } else {
+        out_[index + pcs.outOff] = in_[index + pcs.inOff];
+    }
+}

kompute/op_gelu.comp

@@ -0,0 +1,142 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const float x = in_[i + pcs.inOff];
+
+    out_[i + pcs.outOff] = 0.5*x*(1.0 + tanh(SQRT_2_OVER_PI*x*(1.0 + GELU_COEF_A*x*x)));
+}

kompute/op_getrows_f16.comp

@@ -0,0 +1,150 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    for (int j = 0; j < pcs.ne00; j++) {
+        out_[i*pcs.nb1 + j + pcs.outOff] = inA[r*pcs.nb01/2+j + pcs.inAOff];
+    }
+}

kompute/op_getrows_q4_0.comp

@@ -0,0 +1,179 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+#define UNALIGNED_INPUT inA
+
+block_q4_0 get_unaligned_block_q4_0(uint index) {
+    block_q4_0 fres;
+    fres.d = u8BufToFloat16(UNALIGNED_INPUT, index);
+    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
+        fres.qs[it] = UNALIGNED_INPUT[index+2+it];
+    }
+    return fres;
+}
+
+void dequantize_row_q4_0(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
+    const uint qk = QK4_0;
+
+    const uint nb = k / qk;
+
+    for (uint i = 0; i < nb; i++) {
+        const block_q4_0 block = get_unaligned_block_q4_0(x + i*sizeof_block_q4_0);
+
+        const float16_t d = block.d;
+
+        for (uint j = 0; j < qk/2; ++j) {
+            const int x0 = (block.qs[j] & 0x0F) - 8;
+            const int x1 = (block.qs[j] >>   4) - 8;
+
+            out_[y+i*qk + j + 0   ] = float(x0)*d;
+            out_[y+i*qk + j + qk/2] = float(x1)*d;
+        }
+    }
+}
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    dequantize_row_q4_0(uint(r*pcs.nb01) + pcs.inAOff, uint(i*pcs.nb1/4) + pcs.outOff, pcs.ne00);
+}

kompute/op_getrows_q4_1.comp

@@ -0,0 +1,181 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+#define UNALIGNED_INPUT inA
+
+block_q4_1 get_unaligned_block_q4_1(uint index) {
+    block_q4_1 fres;
+    fres.d = u8BufToFloat16(UNALIGNED_INPUT, index);
+    fres.m = u8BufToFloat16(UNALIGNED_INPUT, index+2);
+    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
+        fres.qs[it] = UNALIGNED_INPUT[index+4+it];
+    }
+    return fres;
+}
+
+void dequantize_row_q4_1(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
+    const uint qk = QK4_1;
+
+    const uint nb = k / qk;
+
+    for (uint i = 0; i < nb; i++) {
+        const block_q4_1 block = get_unaligned_block_q4_1(x + i*sizeof_block_q4_0);
+
+        const float16_t d = block.d;
+        const float16_t m = block.m;
+
+        for (uint j = 0; j < qk/2; ++j) {
+            const int x0 = (block.qs[j] & 0x0F);
+            const int x1 = (block.qs[j] >>   4);
+
+            out_[y+i*qk + j + 0   ] = float(x0)*d + m;
+            out_[y+i*qk + j + qk/2] = float(x1)*d + m;
+        }
+    }
+}
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    dequantize_row_q4_1(uint(r*pcs.nb01) + pcs.inAOff, uint(i*pcs.nb1/4) + pcs.outOff, pcs.ne00);
+}

kompute/op_mul.comp

@@ -0,0 +1,145 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    uint row;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+
+    out_[i + pcs.outOff] = inA[i + pcs.inAOff] * inB[(i) + pcs.inBOff];
+}

kompute/op_mul_mat_f16.comp

@@ -0,0 +1,177 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 64) in;
+
+layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    uint nb01;
+    uint nb02;
+    uint nb11;
+    uint nb12;
+    int ne0;
+    int ne1;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x];
+
+void main() {
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint x = (r0*pcs.nb01 + im*pcs.nb02) / 2 + pcs.inAOff; // Based from inA
+    const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // based from inB
+
+    sum[gl_LocalInvocationID.x] = 0.0;
+
+    for (uint i = gl_LocalInvocationID.x; i < pcs.ne00; i += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += float(inA[x+i]) * float(inB[y+i]);
+    }
+
+    // accumulate the sum from all threads in the threadgroup
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    if (gl_LocalInvocationID.x == 0) {
+        out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = sum[0];
+    }
+}

kompute/op_mul_mat_q4_0.comp

@@ -0,0 +1,195 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne10;
+    int ne0;
+} pcs;
+
+shared float sum[64];
+
+void main() {
+    const uint nb = uint(pcs.ne00/QK4_0);
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+
+    const uint x = r0*nb; // Based from inA without base offset
+    const uint y = r1*uint(pcs.ne10) + pcs.inBOff; // Based from inB
+
+    const uint nth = gl_WorkGroupSize.x*gl_WorkGroupSize.y;
+    const uint ith = gl_WorkGroupSize.y*gl_LocalInvocationID.x + gl_LocalInvocationID.y;
+
+    const uint ix = gl_LocalInvocationID.y/4;           // 0 or 1
+    const uint iy = gl_LocalInvocationID.y - 4*ix;      // 0...3
+
+    const uint first = 4 * iy;
+
+    float sumf = 0.0;
+
+    for (uint i = 2*gl_LocalInvocationID.x + ix; i < nb; i += 2*gl_WorkGroupSize.x) {
+        const uint index = (x+i)*sizeof_block_q4_0+pcs.inAOff;
+        const float d = float(u8BufToFloat16(inA, index));
+
+        const uint xl = first; // Based from bl->qs
+        const uint yl = y + i * QK4_0 + first; // Based from inB
+
+        vec2 acc = vec2(0.0, 0.0);
+
+        for (int j = 0; j < 4; ++j) {
+            const uint8_t b = inA[index+2+xl+j];
+            acc.x += inB[yl+j] * (b & 0xF) + inB[yl+j+16] * (b >> 4);
+            acc.y += inB[yl+j] + inB[yl+j+16];
+        }
+
+        sumf += d * (acc.x - 8.*acc.y);
+    }
+
+    sum[ith] = sumf;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    //
+    barrier();
+    if (ith == 0) {
+        float sumTotal = 0.0;
+        for (uint i = 0; i < nth; ++i) {
+            sumTotal += sum[i];
+        }
+        out_[r1*uint(pcs.ne0) + r0 + pcs.outOff] = sumTotal;
+    }
+}

kompute/op_mul_mat_q4_1.comp

@@ -0,0 +1,218 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne10;
+    int ne0;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x*gl_WorkGroupSize.y];
+
+#define UNALIGNED_INPUT inA
+
+block_q4_1 get_unaligned_block_q4_1(uint index) {
+    block_q4_1 fres;
+    fres.d = u8BufToFloat16(UNALIGNED_INPUT, index);
+    fres.m = u8BufToFloat16(UNALIGNED_INPUT, index+2);
+    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
+        fres.qs[it] = UNALIGNED_INPUT[index+4+it];
+    }
+    return fres;
+}
+
+void main() {
+    const uint nb = uint(pcs.ne00/QK4_1);
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+
+    const uint x = r0*nb; // Based from inA without base offset
+    const uint y = r1*uint(pcs.ne10) + pcs.inBOff; // Based from inB
+
+    const uint nth = gl_WorkGroupSize.x*gl_WorkGroupSize.y;
+    const uint ith = gl_WorkGroupSize.y*gl_LocalInvocationID.x + gl_LocalInvocationID.y;
+
+    const uint ix = gl_LocalInvocationID.y/4;           // 0 or 1
+    const uint iy = gl_LocalInvocationID.y - 4*ix;      // 0...3
+
+    const uint first = 4 * iy;
+
+    float sumf = 0.0;
+
+    for (uint i = 2*gl_LocalInvocationID.x + ix; i < nb; i += 2*gl_WorkGroupSize.x) {
+        //TODO: Removing the use of pointers has been quite hairy here. If something goes wrong here, this is most likely it:
+
+        const block_q4_1 block = get_unaligned_block_q4_1((x+i)*sizeof_block_q4_1+pcs.inAOff);
+
+        const float d = float(block.d);
+        const float m = float(block.m);
+
+        const uint xl = first; // Based from bl->qs
+        const uint yl = y + i * QK4_1 + first; // Based from inB
+
+        vec2 acc = vec2(0.0, 0.0);
+
+        for (int j = 0; j < 4; ++j) {
+            acc.x += inB[yl+j] * (d * (block.qs[xl+j] & 0xF) + m);
+            acc.y += inB[yl+j+16] * (d * (block.qs[xl+j] >> 4) + m);
+        }
+
+        sumf += d * (acc.x - acc.y);
+    }
+
+    sum[ith] = sumf;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    //
+    barrier();
+    memoryBarrierShared();
+    if (ith%4 == 0) {
+        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
+    }
+    barrier();
+    memoryBarrierShared();
+    if (ith%16 == 0) {
+        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
+    }
+    barrier();
+    memoryBarrierShared();
+    if (ith == 0) {
+        for (uint i = 16; i < nth; i += 16) sum[0] += sum[i];
+        out_[r1*uint(pcs.ne0) + r0 + pcs.outOff] = sum[0];
+    }
+}

kompute/op_mulrow.comp

@@ -0,0 +1,145 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    uint row;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+
+    out_[i + pcs.outOff] = inA[i + pcs.inAOff] * inB[(i % pcs.row) + pcs.inBOff];
+}

kompute/op_norm.comp

@@ -0,0 +1,209 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+#define nth 256
+
+layout(local_size_x = nth) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[nth];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+    // MEAN
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = nth/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float mean = sum[0];
+
+    // recenter
+    const uint y = (gl_WorkGroupID.x*pcs.ne00/4) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        out_[y+i00] = in_[x+i00] - mean;
+    }
+
+    // VARIANCE
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = nth/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float variance = sum[0];
+
+    const float scale = 1.0f/sqrt(variance + pcs.eps);
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        out_[y+i00] *= scale;
+    }
+}

kompute/op_relu.comp

@@ -0,0 +1,141 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+
+    out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
+}

kompute/op_rmsnorm.comp

@@ -0,0 +1,178 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+#define nth 256
+
+layout(local_size_x = nth) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[nth];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = nth/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+
+    const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
+
+    const uint y = (gl_WorkGroupID.x*pcs.ne00/4) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        out_[y+i00] = in_[x+i00] * scale;
+    }
+}

kompute/op_rope.comp

@@ -0,0 +1,183 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorIn { float in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    uint n_past;
+    int n_dims;
+    int mode;
+    float freq_base;
+    float freq_scale;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    const bool is_neox = (pcs.mode & 2) != 0;
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    const uint p = ((pcs.mode & 1) == 0 ? pcs.n_past + i2 : i2);
+
+    float theta = pcs.freq_scale * float(p);
+
+    if (!is_neox) {
+        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
+            const float cos_theta = cos(theta);
+            const float sin_theta = sin(theta);
+
+            theta *= theta_scale;
+
+            const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = in_[src];
+            const float x1 = in_[src+1];
+
+            out_[dst_data] = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
+        }
+    } else {
+        // TODO: implement
+    }
+}

kompute/op_scale.comp

@@ -0,0 +1,142 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    float scale;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+
+    out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
+}

kompute/op_silu.comp

@@ -0,0 +1,141 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const float x = in_[i + pcs.inOff];
+
+    out_[i + pcs.outOff] = x / (1.0 + exp(-x));
+}

kompute/op_softmax.comp

@@ -0,0 +1,197 @@
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#version 450
+
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+#define BM 128
+#define BN 128
+#define BK 8
+#define TM 8
+#define TN 8
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+#define sizeof_block_q4_1 0x14
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+
+#ifndef QK_K
+#define QK_K 256
+#endif
+
+#if QK_K == 256
+#define K_SCALE_SIZE 12
+#else
+#define K_SCALE_SIZE 4
+#endif
+
+struct block_q2_K {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    float16_t d;           // super-block scale for quantized scales
+    float16_t dmin;        // super-block scale for quantized mins
+};
+// 84 bytes / block
+
+struct block_q3_K {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+#if QK_K == 64
+    uint8_t scales[2];
+#else
+    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
+#endif
+    float16_t d;             // super-block scale
+};
+
+#if QK_K == 64
+typedef struct {
+    float16_t    d[2];          // super-block scales/mins
+    uint8_t scales[2];
+    uint8_t qs[QK_K/2];    // 4-bit quants
+} block_q4_K;
+#else
+struct block_q4_K {
+    float16_t d;             // super-block scale for quantized scales
+    float16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+};
+#endif
+
+#if QK_K == 64
+struct block_q5_K {
+    float16_t  d;                     // super-block scales/mins
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+#else
+struct block_q5_K {
+    float16_t d;                      // super-block scale for quantized scales
+    float16_t dmin;                   // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+};
+// 176 bytes / block
+#endif
+
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;                  // super-block scale
+};
+// 210 bytes / block
+
+#define nth 32
+
+layout(local_size_x = nth) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+} pcs;
+
+shared float buf[nth];
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
+    const uint psrc0 = extra_off + pcs.inOff; // Based from in_
+    const uint pdst = extra_off + pcs.outOff; // Based from out_
+
+    // parallel max
+    buf[gl_LocalInvocationID.x] = uintBitsToFloat(0xFF800000);
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        buf[gl_LocalInvocationID.x] = max(buf[gl_LocalInvocationID.x], in_[psrc0 + i00]);
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = nth/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            buf[gl_LocalInvocationID.x] = max(buf[gl_LocalInvocationID.x], buf[gl_LocalInvocationID.x + i]);
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    const float max_ = buf[0];
+
+    // parallel sum
+    buf[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        buf[gl_LocalInvocationID.x] += exp(in_[psrc0 + i00] - max_);
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = nth/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            buf[gl_LocalInvocationID.x] += buf[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    const float sum = buf[0];
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += nth) {
+        out_[pdst + i00] = exp(in_[psrc0 + i00] - max_) / sum;
+    }
+}

kompute/scripts/convert_shaders.py

@@ -0,0 +1,148 @@
+"""
+    Script to handle conversion of compute shaders to spirv and to headers
+"""
+import os
+import sys
+import logging
+import click
+import subprocess
+
+logger = logging.getLogger(__name__)
+logger.addHandler(logging.StreamHandler())
+
+is_windows = sys.platform.startswith('win')
+
+CWD=os.path.dirname(os.path.abspath(__file__))
+XXD_LINUX_CMD="xxd"
+XXD_WINDOWS_CMD=os.path.abspath(os.path.join(CWD, "..\\external\\bin\\", "xxd.exe"))
+
+SHADER_GENERATED_NOTICE = """/*
+    THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT
+
+    ---
+
+    Copyright 2020 The Institute for Ethical AI & Machine Learning
+
+    Licensed under the Apache License, Version 2.0 (the "License");
+    you may not use this file except in compliance with the License.
+    You may obtain a copy of the License at
+
+        http://www.apache.org/licenses/LICENSE-2.0
+
+    Unless required by applicable law or agreed to in writing, software
+    distributed under the License is distributed on an "AS IS" BASIS,
+    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+    See the License for the specific language governing permissions and
+    limitations under the License.
+*/
+"""
+
+@click.command()
+@click.option(
+    "--shader-path",
+    "-p",
+    envvar="KOMPUTE_SHADER_PATH",
+    required=True,
+    help="The path for the directory to build and convert shaders",
+)
+@click.option(
+    "--shader-binary",
+    "-s",
+    envvar="KOMPUTE_SHADER_BINARY",
+    required=True,
+    help="The path for the directory to build and convert shaders",
+)
+@click.option(
+    "--header-path",
+    "-c",
+    envvar="KOMPUTE_HEADER_PATH",
+    default="",
+    required=False,
+    help="The (optional) output file for the cpp header files",
+)
+@click.option(
+    "--verbose",
+    "-v",
+    envvar="KOMPUTE_HEADER_PATH",
+    default=False,
+    is_flag=True,
+    help="Enable versbosity if flag is provided",
+)
+def run_cli(
+    shader_path: str = None,
+    shader_binary: str = None,
+    header_path: bool = None,
+    verbose: bool = None,
+):
+    """
+    CLI function for shader generation
+    """
+
+    if verbose:
+        logger.setLevel(logging.DEBUG)
+    else:
+        logger.setLevel(logging.WARNING)
+
+    logger.debug(f"Starting script with variables: {locals()}")
+
+    if is_windows:
+        logger.debug(f"Running on windows, converting input paths")
+        shader_path = shader_path.replace("/", "\\")
+        header_path = header_path.replace("/", "\\")
+
+    shader_files = []
+    for root, directory, files in os.walk(shader_path):
+        for file in files:
+            if file.endswith(".comp"):
+                shader_files.append(os.path.join(root, file))
+
+    run_cmd = lambda *args: subprocess.check_output([*args]).decode()
+
+    logger.debug(f"Output spirv path: {shader_path}")
+    logger.debug(f"Converting files to spirv: {shader_files}")
+
+    spirv_files = []
+    for file in shader_files:
+        logger.debug(f"Converting to spirv: {file}")
+        spirv_file = f"{file}.spv"
+        run_cmd(shader_binary, "-V", file, "-o", spirv_file)
+        spirv_files.append(spirv_file)
+
+    # Create cpp files if header_path provided
+    if header_path:
+        logger.debug(f"Header path provided. Converting bin files to hpp.")
+        logger.debug(f"Output header path: {shader_path}")
+
+        # Check if xxd command options are available
+        if is_windows:
+            xxd_cmd = XXD_WINDOWS_CMD
+        else:
+            xxd_cmd = XXD_LINUX_CMD
+
+        for file in spirv_files:
+            print(xxd_cmd)
+            header_data = str(run_cmd(xxd_cmd, "-i", file))
+            # Ensuring the variable is a static const unsigned
+            header_data = header_data.replace("unsigned", "static const unsigned")
+            if is_windows:
+                raw_file_name = file.split("\\")[-1]
+            else:
+                raw_file_name = file.split("/")[-1]
+            file_name = f"shader{raw_file_name}"
+            header_file = file_name.replace(".comp.spv", ".hpp")
+            header_file_define = "SHADEROP_" + header_file.replace(".", "_").upper()
+            logger.debug(f"Converting to hpp: {file_name}")
+            with open(os.path.join(header_path, header_file), "w+", newline='\n') as fstream:
+                fstream.write(f"{SHADER_GENERATED_NOTICE}\n")
+                fstream.write(f"#ifndef {header_file_define}\n")
+                fstream.write(f"#define {header_file_define}\n\n")
+                fstream.write("namespace kp {\n")
+                fstream.write("namespace shader_data {\n")
+                fstream.write(f"{header_data}")
+                fstream.write("}\n")
+                fstream.write("}\n")
+                fstream.write(f"#endif // define {header_file_define}\n")
+
+
+if __name__ == "__main__":
+    run_cli()

kompute/scripts/requirements.txt

@@ -0,0 +1,11 @@
+# CLI dependencies
+click==7.1.2
+
+# Dev dependencies
+black==19.10b0
+quom==1.2.0
+Sphinx==3.2.1
+sphinx_material==0.0.30
+breathe==4.20.0
+m2r2==0.2.5
+git+git://github.com/pybind/pybind11_mkdoc.git@master

kompute/setup.py

@@ -0,0 +1,93 @@
+import os
+import re
+import platform
+import sys
+import sysconfig
+import subprocess
+
+from setuptools import setup, Extension
+from setuptools.command.build_ext import build_ext
+from distutils.version import LooseVersion
+
+curr_dir = os.path.abspath(os.path.dirname(__file__))
+with open(os.path.join(curr_dir, 'README.md'), encoding='utf-8') as f:
+    long_description = f.read()
+
+class CMakeExtension(Extension):
+    def __init__(self, name, sourcedir=''):
+        Extension.__init__(self, name, sources=[])
+        self.sourcedir = os.path.abspath(sourcedir)
+
+
+class CMakeBuild(build_ext):
+    def run(self):
+        try:
+            out = subprocess.check_output(['cmake', '--version'])
+        except OSError:
+            raise RuntimeError("CMake must be installed to build the following extensions: " +
+                               ", ".join(e.name for e in self.extensions))
+
+        cmake_version = LooseVersion(re.search(r'version\s*([\d.]+)', out.decode()).group(1))
+        if cmake_version < '3.15':
+            raise RuntimeError("CMake >= 3.15 is required")
+
+        for ext in self.extensions:
+            self.build_extension(ext)
+
+    def build_extension(self, ext):
+        extdir = os.path.abspath(os.path.dirname(self.get_ext_fullpath(ext.name)))
+        # required for auto-detection of auxiliary "native" libs
+        if not extdir.endswith(os.path.sep):
+            extdir += os.path.sep
+
+        cmake_args = ['-DCMAKE_LIBRARY_OUTPUT_DIRECTORY=' + extdir,
+                      '-DKOMPUTE_OPT_BUILD_PYTHON=ON',
+                      '-DKOMPUTE_OPT_LOG_LEVEL=Off',
+                      '-DKOMPUTE_OPT_USE_SPDLOG=Off',
+                      '-DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
+                      '-DPYTHON_EXECUTABLE=' + sys.executable,
+                      '-DPYTHON_INCLUDE_DIR=' + sysconfig.get_path('include'),
+                      '-DPYTHON_LIBRARY=' + sysconfig.get_path('stdlib'),
+        ]
+
+        cfg = 'Debug' if self.debug else 'Release'
+        build_args = ['--config', cfg]
+
+        env = os.environ.copy()
+        oldCxxFlags = env.get('CXXFLAGS', '')
+        env['CXXFLAGS'] = f'{oldCxxFlags} -DVERSION_INFO=\\"{self.distribution.get_version()}\\"'
+
+        if platform.system() == "Windows":
+            cmake_args += [f'-DCMAKE_LIBRARY_OUTPUT_DIRECTORY_{cfg.upper()}={extdir}']
+            if sys.maxsize > 2**32:
+                cmake_args += ['-A', 'x64']
+            build_args += ['--', '/m']
+        else:
+            env['CXXFLAGS'] += ' -fPIC'
+            cmake_args += ['-DCMAKE_BUILD_TYPE=' + cfg]
+            build_args += ['--', '-j']
+            # Optional environment variable to limit the number of parallel jobs for GitHub actions to reduce RAM usage
+            if 'KOMPUTE_PYTHON_NUM_PARALLEL_THREADS' in env:
+                build_args += env['KOMPUTE_PYTHON_NUM_PARALLEL_THREADS']
+
+        if not os.path.exists(self.build_temp):
+            os.makedirs(self.build_temp)
+
+        subprocess.check_call(['cmake', ext.sourcedir] + cmake_args, cwd=self.build_temp, env=env)
+        subprocess.check_call(['cmake', '--build', '.'] + build_args, cwd=self.build_temp)
+
+setup(
+    name='kp',
+    version='0.8.1',
+    author='Alejandro Saucedo',
+    description='Kompute: Blazing fast, mobile-enabled, asynchronous, and optimized for advanced GPU processing usecases.',
+    long_description=long_description,
+    long_description_content_type='text/markdown',
+    ext_modules=[CMakeExtension('kp')],
+    install_requires=[
+        "numpy<2.0.0"
+    ],
+    cmdclass=dict(build_ext=CMakeBuild),
+    zip_safe=False,
+    include_package_data=True,
+)

kompute/src/Algorithm.cpp

@@ -0,0 +1,428 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include <fstream>
+
+#include "kompute/Algorithm.hpp"
+
+namespace kp {
+
+Algorithm::~Algorithm()
+{
+    KP_LOG_DEBUG("Kompute Algorithm Destructor started");
+
+    this->destroy();
+}
+
+bool
+Algorithm::isInit()
+{
+    return this->mPipeline && this->mPipelineCache && this->mPipelineLayout &&
+           this->mDescriptorPool && this->mDescriptorSet &&
+           this->mDescriptorSetLayout && this->mShaderModule;
+}
+
+void
+Algorithm::destroy()
+{
+    // We don't have to free memory on destroy as it's freed by the
+    // commandBuffer destructor if (this->mPushConstantsData) {
+    //     free(this->mPushConstantsData);
+    // }
+    // if (this->mSpecializationConstantsData) {
+    //     free(this->mSpecializationConstantsData);
+    // }
+
+    if (!this->mDevice) {
+        KP_LOG_WARN("Kompute Algorithm destroy function reached with null "
+                    "Device pointer");
+        return;
+    }
+
+    if (this->mFreePipeline && this->mPipeline) {
+        KP_LOG_DEBUG("Kompute Algorithm Destroying pipeline");
+        if (!this->mPipeline) {
+            KP_LOG_WARN("Kompute Algorithm Error requested to destroy "
+                        "pipeline but it is null");
+        }
+        this->mDevice->destroy(
+          *this->mPipeline,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        this->mPipeline = nullptr;
+    }
+
+    if (this->mFreePipelineLayout && this->mPipelineLayout) {
+        KP_LOG_DEBUG("Kompute Algorithm Destroying pipeline layout");
+        if (!this->mPipelineLayout) {
+            KP_LOG_WARN("Kompute Algorithm Error requested to destroy "
+                        "pipeline layout but it is null");
+        }
+        this->mDevice->destroy(
+          *this->mPipelineLayout,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        this->mPipelineLayout = nullptr;
+    }
+
+    if (this->mFreeShaderModule && this->mShaderModule) {
+        KP_LOG_DEBUG("Kompute Algorithm Destroying shader module");
+        if (!this->mShaderModule) {
+            KP_LOG_WARN("Kompute Algorithm Error requested to destroy shader "
+                        "module but it is null");
+        }
+        this->mDevice->destroy(
+          *this->mShaderModule,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        this->mShaderModule = nullptr;
+    }
+
+    freeParameters();
+}
+
+void
+Algorithm::freeParameters()
+{
+    if (this->mFreeDescriptorSetLayout && this->mDescriptorSetLayout) {
+        KP_LOG_DEBUG("Kompute Algorithm Destroying Descriptor Set Layout");
+        if (!this->mDescriptorSetLayout) {
+            KP_LOG_WARN("Kompute Algorithm Error requested to destroy "
+                        "descriptor set layout but it is null");
+        }
+        this->mDevice->destroy(
+          *this->mDescriptorSetLayout,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        this->mDescriptorSetLayout = nullptr;
+    }
+}
+
+void
+Algorithm::createParameters()
+{
+    KP_LOG_DEBUG("Kompute Algorithm createParameters started");
+    if (!*this->mDescriptorPool) {
+        KP_LOG_ERROR("Kompute Algorithm can not create descriptor pool");
+        return;
+    }
+
+    std::vector<vk::DescriptorSetLayoutBinding> descriptorSetBindings;
+    for (size_t i = 0; i < this->mTensors.size(); i++) {
+        descriptorSetBindings.push_back(
+          vk::DescriptorSetLayoutBinding(i, // Binding index
+                                         vk::DescriptorType::eStorageBuffer,
+                                         1, // Descriptor count
+                                         vk::ShaderStageFlagBits::eCompute));
+    }
+
+    // This is the component that is fed into the pipeline
+    vk::DescriptorSetLayoutCreateInfo descriptorSetLayoutInfo(
+      vk::DescriptorSetLayoutCreateFlags(),
+      static_cast<uint32_t>(descriptorSetBindings.size()),
+      descriptorSetBindings.data());
+
+    KP_LOG_DEBUG("Kompute Algorithm creating descriptor set layout");
+    this->mDescriptorSetLayout = std::make_shared<vk::DescriptorSetLayout>();
+    vk::Result result = this->mDevice->createDescriptorSetLayout(
+      &descriptorSetLayoutInfo, nullptr, this->mDescriptorSetLayout.get());
+
+   if (result != vk::Result::eSuccess) {
+        KP_LOG_ERROR("Failed to create descriptor set layout. Error code: {}", vk::to_string(result));
+    } else {
+        this->mFreeDescriptorSetLayout = true;
+        KP_LOG_DEBUG("Successfully allocated descriptor set layout.");
+    }
+
+    vk::DescriptorSetAllocateInfo descriptorSetAllocateInfo(
+      *this->mDescriptorPool,
+      1, // Descriptor set layout count
+      this->mDescriptorSetLayout.get());
+
+    KP_LOG_DEBUG("Kompute Algorithm allocating descriptor sets");
+    this->mDescriptorSet = std::make_shared<vk::DescriptorSet>();
+    result = this->mDevice->allocateDescriptorSets(&descriptorSetAllocateInfo,
+                                          this->mDescriptorSet.get());
+
+    if (result != vk::Result::eSuccess) {
+        KP_LOG_ERROR("Failed to allocate descriptor sets. Error code: {}", vk::to_string(result));
+    } else {
+        this->mFreeDescriptorSet = true;
+        KP_LOG_DEBUG("Successfully allocated descriptor sets.");
+    }
+
+    this->mFreeDescriptorSet = true;
+
+    KP_LOG_DEBUG("Kompute Algorithm updating descriptor sets");
+    for (size_t i = 0; i < this->mTensors.size(); i++) {
+        std::vector<vk::WriteDescriptorSet> computeWriteDescriptorSets;
+
+        vk::DescriptorBufferInfo descriptorBufferInfo =
+          this->mTensors[i]->constructDescriptorBufferInfo();
+
+        computeWriteDescriptorSets.push_back(
+          vk::WriteDescriptorSet(*this->mDescriptorSet,
+                                 i, // Destination binding
+                                 0, // Destination array element
+                                 1, // Descriptor count
+                                 vk::DescriptorType::eStorageBuffer,
+                                 nullptr, // Descriptor image info
+                                 &descriptorBufferInfo));
+
+        this->mDevice->updateDescriptorSets(computeWriteDescriptorSets,
+                                            nullptr);
+    }
+
+    KP_LOG_DEBUG("Kompute Algorithm successfully run init");
+}
+
+void
+Algorithm::updateParameters()
+{
+    KP_LOG_DEBUG("Kompute Algorithm updateParameters started");
+    if (!*this->mDescriptorPool) {
+        KP_LOG_ERROR("Kompute Algorithm can not create descriptor pool");
+        return;
+    }
+
+    vk::DescriptorSetAllocateInfo descriptorSetAllocateInfo(
+      *this->mDescriptorPool,
+      1, // Descriptor set layout count
+      this->mDescriptorSetLayout.get());
+
+    KP_LOG_DEBUG("Kompute Algorithm allocating descriptor sets");
+    this->mDescriptorSet = std::make_shared<vk::DescriptorSet>();
+    vk::Result result = this->mDevice->allocateDescriptorSets(&descriptorSetAllocateInfo,
+                                          this->mDescriptorSet.get());
+
+    if (result != vk::Result::eSuccess) {
+        KP_LOG_ERROR("Failed to allocate descriptor sets. Error code: {}", vk::to_string(result));
+    } else {
+        this->mFreeDescriptorSet = true;
+        KP_LOG_DEBUG("Successfully allocated descriptor sets.");
+    }
+
+    this->mFreeDescriptorSet = true;
+
+    KP_LOG_DEBUG("Kompute Algorithm updating descriptor sets");
+    for (size_t i = 0; i < this->mTensors.size(); i++) {
+        std::vector<vk::WriteDescriptorSet> computeWriteDescriptorSets;
+
+        vk::DescriptorBufferInfo descriptorBufferInfo =
+          this->mTensors[i]->constructDescriptorBufferInfo();
+
+        computeWriteDescriptorSets.push_back(
+          vk::WriteDescriptorSet(*this->mDescriptorSet,
+                                 i, // Destination binding
+                                 0, // Destination array element
+                                 1, // Descriptor count
+                                 vk::DescriptorType::eStorageBuffer,
+                                 nullptr, // Descriptor image info
+                                 &descriptorBufferInfo));
+
+        this->mDevice->updateDescriptorSets(computeWriteDescriptorSets,
+                                            nullptr);
+    }
+
+    KP_LOG_DEBUG("Kompute Algorithm successfully run init");
+}
+
+void
+Algorithm::createShaderModule()
+{
+    KP_LOG_DEBUG("Kompute Algorithm createShaderModule started");
+
+    vk::ShaderModuleCreateInfo shaderModuleInfo(vk::ShaderModuleCreateFlags(),
+                                                sizeof(uint32_t) *
+                                                  this->mSpirv.size(),
+                                                this->mSpirv.data());
+
+    KP_LOG_DEBUG("Kompute Algorithm Creating shader module. ShaderFileSize: {}",
+                 this->mSpirv.size());
+    this->mFreeShaderModule = true;
+    this->mShaderModule = std::make_shared<vk::ShaderModule>();
+    this->mDevice->createShaderModule(
+      &shaderModuleInfo, nullptr, this->mShaderModule.get());
+    this->mFreeShaderModule = true;
+
+    KP_LOG_DEBUG("Kompute Algorithm create shader module success");
+}
+
+void
+Algorithm::createPipeline()
+{
+    KP_LOG_DEBUG("Kompute Algorithm calling create Pipeline");
+
+    vk::PipelineLayoutCreateInfo pipelineLayoutInfo(
+      vk::PipelineLayoutCreateFlags(),
+      1, // Set layout count
+      this->mDescriptorSetLayout.get());
+
+    vk::PushConstantRange pushConstantRange;
+    if (this->mPushConstantsSize) {
+        pushConstantRange.setStageFlags(vk::ShaderStageFlagBits::eCompute);
+        pushConstantRange.setOffset(0);
+        pushConstantRange.setSize(this->mPushConstantsDataTypeMemorySize *
+                                  this->mPushConstantsSize);
+
+        pipelineLayoutInfo.setPushConstantRangeCount(1);
+        pipelineLayoutInfo.setPPushConstantRanges(&pushConstantRange);
+    }
+
+    this->mPipelineLayout = std::make_shared<vk::PipelineLayout>();
+    this->mDevice->createPipelineLayout(
+      &pipelineLayoutInfo, nullptr, this->mPipelineLayout.get());
+    this->mFreePipelineLayout = true;
+
+    std::vector<vk::SpecializationMapEntry> specializationEntries;
+
+    for (uint32_t i = 0; i < this->mSpecializationConstantsSize; i++) {
+        vk::SpecializationMapEntry specializationEntry(
+          static_cast<uint32_t>(i),
+          static_cast<uint32_t>(
+            this->mSpecializationConstantsDataTypeMemorySize * i),
+          this->mSpecializationConstantsDataTypeMemorySize);
+
+        specializationEntries.push_back(specializationEntry);
+    }
+
+    // This passes ownership of the memory so we remove ownership from
+    // specialization container by using "transferDataOwnership"
+    vk::SpecializationInfo specializationInfo(
+      static_cast<uint32_t>(specializationEntries.size()),
+      specializationEntries.data(),
+      this->mSpecializationConstantsDataTypeMemorySize *
+        this->mSpecializationConstantsSize,
+      this->mSpecializationConstantsData);
+
+    vk::PipelineShaderStageCreateInfo shaderStage(
+      vk::PipelineShaderStageCreateFlags(),
+      vk::ShaderStageFlagBits::eCompute,
+      *this->mShaderModule,
+      "main",
+      &specializationInfo);
+
+    vk::ComputePipelineCreateInfo pipelineInfo(vk::PipelineCreateFlags(),
+                                               shaderStage,
+                                               *this->mPipelineLayout,
+                                               vk::Pipeline(),
+                                               0);
+
+#ifdef KOMPUTE_CREATE_PIPELINE_RESULT_VALUE
+    vk::ResultValue<vk::Pipeline> pipelineResult =
+      this->mDevice->createComputePipeline(*mPipelineCache, pipelineInfo);
+
+    if (pipelineResult.result != vk::Result::eSuccess) {
+        throw std::runtime_error("Failed to create pipeline result: " +
+                                 vk::to_string(pipelineResult.result));
+    }
+
+    vk::Pipeline& pipeline = pipelineResult.value;
+    this->mPipeline = std::make_shared<vk::Pipeline>(pipeline);
+    this->mFreePipeline = true;
+#else
+    vk::Pipeline pipeline =
+      this->mDevice->createComputePipeline(*mPipelineCache, pipelineInfo)
+        .value;
+    this->mPipeline = std::make_shared<vk::Pipeline>(pipeline);
+    this->mFreePipeline = true;
+#endif
+
+    // TODO: Update to consistent
+    // this->mPipeline = std::make_shared<vk::Pipeline>();
+    // this->mDevice->createComputePipelines(
+    //         *this->mPipelineCache, 1, &pipelineInfo, nullptr,
+    //         this->mPipeline.get());
+
+    KP_LOG_DEBUG("Kompute Algorithm Create Pipeline Success");
+}
+
+void
+Algorithm::recordBindCore(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute Algorithm binding pipeline");
+
+    commandBuffer.bindPipeline(vk::PipelineBindPoint::eCompute,
+                               *this->mPipeline);
+
+    KP_LOG_DEBUG("Kompute Algorithm binding descriptor sets");
+
+    commandBuffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
+                                     *this->mPipelineLayout,
+                                     0, // First set
+                                     *this->mDescriptorSet,
+                                     nullptr // Dispatcher
+    );
+}
+
+void
+Algorithm::recordBindPush(const vk::CommandBuffer& commandBuffer)
+{
+    if (this->mPushConstantsSize) {
+        KP_LOG_DEBUG("Kompute Algorithm binding push constants memory size: {}",
+                     this->mPushConstantsSize *
+                       this->mPushConstantsDataTypeMemorySize);
+
+        commandBuffer.pushConstants(*this->mPipelineLayout,
+                                    vk::ShaderStageFlagBits::eCompute,
+                                    0,
+                                    this->mPushConstantsSize *
+                                      this->mPushConstantsDataTypeMemorySize,
+                                    this->mPushConstantsData);
+    }
+}
+
+void
+Algorithm::recordDispatch(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute Algorithm recording dispatch");
+
+    commandBuffer.dispatch(
+      this->mWorkgroup[0], this->mWorkgroup[1], this->mWorkgroup[2]);
+}
+
+void
+Algorithm::setWorkgroup(const Workgroup& workgroup, uint32_t minSize)
+{
+
+    KP_LOG_INFO("Kompute OpAlgoCreate setting dispatch size");
+
+    // The dispatch size is set up based on either explicitly provided template
+    // parameters or by default it would take the shape and size of the tensors
+    if (workgroup[0] > 0) {
+        // If at least the x value is provided we use mainly the parameters
+        // provided
+        this->mWorkgroup = { workgroup[0],
+                             workgroup[1] > 0 ? workgroup[1] : 1,
+                             workgroup[2] > 0 ? workgroup[2] : 1 };
+    } else {
+        this->mWorkgroup = { minSize, 1, 1 };
+    }
+
+    KP_LOG_INFO("Kompute OpAlgoCreate set dispatch size X: {}, Y: {}, Z: {}",
+                this->mWorkgroup[0],
+                this->mWorkgroup[1],
+                this->mWorkgroup[2]);
+}
+
+const Workgroup&
+Algorithm::getWorkgroup()
+{
+    return this->mWorkgroup;
+}
+
+const std::vector<std::shared_ptr<Tensor>>&
+Algorithm::getTensors()
+{
+    return this->mTensors;
+}
+
+void Algorithm::setTensors(const std::vector<std::shared_ptr<Tensor>>& tensors)
+{
+    this->mTensors = tensors;
+}
+
+}

kompute/src/CMakeLists.txt

@@ -0,0 +1,85 @@
+# SPDX-License-Identifier: Apache-2.0
+
+cmake_minimum_required(VERSION 3.20)
+
+if(KOMPUTE_OPT_ANDROID_BUILD)
+    find_library(android android)
+endif()
+
+cmake_minimum_required(VERSION 3.20)
+
+add_library(kompute STATIC Algorithm.cpp
+    Manager.cpp
+    OpAlgoDispatch.cpp
+    OpMemoryBarrier.cpp
+    OpTensorCopy.cpp
+    OpTensorSyncDevice.cpp
+    OpTensorSyncLocal.cpp
+    OpBufferSyncDevice.cpp
+    OpBufferSyncLocal.cpp
+    Sequence.cpp
+    Tensor.cpp
+    Core.cpp)
+
+add_library(kompute::kompute ALIAS kompute)
+
+# Set version for shared libraries.
+set_target_properties(kompute
+    PROPERTIES
+    VERSION ${${PROJECT_NAME}_VERSION}
+    SOVERSION ${${PROJECT_NAME}_VERSION_MAJOR}
+    POSITION_INDEPENDENT_CODE TRUE)
+
+# Import GNU common install directory variables
+include(GNUInstallDirs)
+
+install(TARGETS kompute
+    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR})
+
+# Include CMake helpers for package config files
+# Follow this installation guideline: https://cmake.org/cmake/help/latest/manual/cmake-packages.7.html
+include(CMakePackageConfigHelpers)
+
+configure_package_config_file(${PROJECT_SOURCE_DIR}/cmake/komputeConfig.cmake.in
+    "${PROJECT_BINARY_DIR}/kompute/komputeConfig.cmake"
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/kompute)
+
+#install(FILES ${PROJECT_BINARY_DIR}/kompute/komputeConfig.cmake
+#    ${PROJECT_BINARY_DIR}/kompute/komputeConfigVersion.cmake DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/kompute)
+
+# ####################################################
+# Linking
+# ####################################################
+if(KOMPUTE_OPT_ANDROID_BUILD)
+    target_link_libraries(kompute PUBLIC vulkanAndroid
+        android
+        kp_logger
+        kp_shader
+        fmt::fmt-header-only)
+else()
+    target_link_libraries(kompute PUBLIC
+        kp_logger
+        kp_shader
+        fmt::fmt-header-only)
+endif()
+
+if(KOMPUTE_OPT_BUILD_PYTHON)
+    include_directories(${PYTHON_INCLUDE_DIRS})
+
+    target_link_libraries(kompute PRIVATE pybind11::headers ${PYTHON_LIBRARIES})
+endif()
+
+if(KOMPUTE_OPT_USE_BUILT_IN_VULKAN_HEADER)
+    target_link_libraries(kompute PUBLIC Vulkan-Headers)
+else()
+    target_link_libraries(kompute PUBLIC Vulkan::Headers)
+endif()
+
+# ####################################################
+# Misc
+# ####################################################
+add_subdirectory(logger)
+add_subdirectory(shaders)
+add_subdirectory(include)

kompute/src/Core.cpp

@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/Core.hpp"
+
+#ifndef KOMPUTE_VK_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE
+#define KOMPUTE_VK_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE
+/**
+ * Ensures support for dynamic loading of Vulkan functions on Android.
+ * Acts as a default store for loaded functions.
+ * More information:
+ * https://github.com/KhronosGroup/Vulkan-Hpp#vulkan_hpp_default_dispatcher
+ **/
+VULKAN_HPP_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE
+#endif // !KOMPUTE_VK_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE
+
+namespace kp {
+} // namespace kp

kompute/src/Manager.cpp

@@ -0,0 +1,513 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/Manager.hpp"
+#include "fmt/format.h"
+#include "kompute/logger/Logger.hpp"
+#include <fmt/core.h>
+#include <iterator>
+#include <set>
+#include <sstream>
+#include <string>
+
+namespace kp {
+
+#ifndef KOMPUTE_DISABLE_VK_DEBUG_LAYERS
+static VKAPI_ATTR VkBool32 VKAPI_CALL
+debugMessageCallback(VkDebugReportFlagsEXT /*flags*/,
+                     VkDebugReportObjectTypeEXT /*objectType*/,
+                     uint64_t /*object*/,
+                     size_t /*location*/,
+                     int32_t /*messageCode*/,
+#if KOMPUTE_OPT_ACTIVE_LOG_LEVEL <= KOMPUTE_LOG_LEVEL_DEBUG
+                     const char* pLayerPrefix,
+                     const char* pMessage,
+#else
+                     const char* /*pLayerPrefix*/,
+                     const char* /*pMessage*/,
+#endif
+                     void* /*pUserData*/)
+{
+    KP_LOG_DEBUG("[VALIDATION]: {} - {}", pLayerPrefix, pMessage);
+    return VK_FALSE;
+}
+#endif
+
+Manager::Manager()
+{
+    this->mManageResources = true;
+
+// Make sure the logger is setup
+#if !KOMPUTE_OPT_LOG_LEVEL_DISABLED
+    logger::setupLogger();
+#endif
+    this->createInstance();
+}
+
+void Manager::initializeDevice(uint32_t physicalDeviceIndex,
+                               const std::vector<uint32_t>& familyQueueIndices,
+                               const std::vector<std::string>& desiredExtensions)
+{
+    this->createDevice(
+      familyQueueIndices, physicalDeviceIndex, desiredExtensions);
+}
+
+Manager::~Manager()
+{
+    KP_LOG_DEBUG("Kompute Manager Destructor started");
+    this->destroy();
+}
+
+void
+Manager::destroy()
+{
+
+    KP_LOG_DEBUG("Kompute Manager destroy() started");
+
+    if (this->mDevice == nullptr) {
+        KP_LOG_ERROR(
+          "Kompute Manager destructor reached with null Device pointer");
+        return;
+    }
+
+    if (this->mManageResources && this->mManagedSequences.size()) {
+        KP_LOG_DEBUG("Kompute Manager explicitly running destructor for "
+                     "managed sequences");
+        for (const std::weak_ptr<Sequence>& weakSq : this->mManagedSequences) {
+            if (std::shared_ptr<Sequence> sq = weakSq.lock()) {
+                sq->destroy();
+            }
+        }
+        this->mManagedSequences.clear();
+    }
+
+    if (this->mManageResources && !this->mManagedAlgorithmsMap.empty()) {
+        KP_LOG_DEBUG("Kompute Manager explicitly freeing algorithms");
+        for (const auto& kv : this->mManagedAlgorithmsMap) {
+            if (std::shared_ptr<Algorithm> algorithm = kv.second) {
+                algorithm->destroy();
+            }
+        }
+        this->mManagedAlgorithmsMap.clear();
+    }
+
+    if (this->mManageResources && this->mManagedTensors.size()) {
+        KP_LOG_DEBUG("Kompute Manager explicitly freeing tensors");
+        for (const std::weak_ptr<Tensor>& weakTensor : this->mManagedTensors) {
+            if (std::shared_ptr<Tensor> tensor = weakTensor.lock()) {
+                tensor->destroy();
+            }
+        }
+        this->mManagedTensors.clear();
+    }
+
+    if (this->mPipelineCache) {
+        KP_LOG_DEBUG("Kompute Manager Destroying pipeline cache");
+        if (!this->mPipelineCache) {
+            KP_LOG_WARN("Kompute Manager Error requested to destroy "
+                        "pipeline cache but it is null");
+        }
+        this->mDevice->destroy(
+          *this->mPipelineCache,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        this->mPipelineCache = nullptr;
+    }
+
+    if (this->mFreeDevice) {
+        KP_LOG_INFO("Destroying device");
+        this->mDevice->destroy(
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        this->mDevice = nullptr;
+        KP_LOG_DEBUG("Kompute Manager Destroyed Device");
+    }
+
+    if (this->mInstance == nullptr) {
+        KP_LOG_ERROR(
+          "Kompute Manager destructor reached with null Instance pointer");
+        return;
+    }
+
+#ifndef KOMPUTE_DISABLE_VK_DEBUG_LAYERS
+    if (this->mDebugReportCallback) {
+        this->mInstance->destroyDebugReportCallbackEXT(
+          this->mDebugReportCallback, nullptr, this->mDebugDispatcher);
+        KP_LOG_DEBUG("Kompute Manager Destroyed Debug Report Callback");
+    }
+#endif
+
+    if (this->mFreeInstance) {
+        this->mInstance->destroy(
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        this->mInstance = nullptr;
+        KP_LOG_DEBUG("Kompute Manager Destroyed Instance");
+    }
+}
+
+void
+Manager::createInstance()
+{
+
+    KP_LOG_DEBUG("Kompute Manager creating instance");
+
+    this->mFreeInstance = true;
+
+    vk::ApplicationInfo applicationInfo;
+    applicationInfo.pApplicationName = "Kompute";
+    applicationInfo.pEngineName = "Kompute";
+    applicationInfo.apiVersion = KOMPUTE_VK_API_VERSION;
+    applicationInfo.engineVersion = KOMPUTE_VK_API_VERSION;
+    applicationInfo.applicationVersion = KOMPUTE_VK_API_VERSION;
+
+    std::vector<const char*> applicationExtensions;
+
+#ifndef KOMPUTE_DISABLE_VK_DEBUG_LAYERS
+    applicationExtensions.push_back(VK_EXT_DEBUG_REPORT_EXTENSION_NAME);
+#endif
+
+    vk::InstanceCreateInfo computeInstanceCreateInfo;
+    computeInstanceCreateInfo.pApplicationInfo = &applicationInfo;
+    if (!applicationExtensions.empty()) {
+        computeInstanceCreateInfo.enabledExtensionCount =
+          (uint32_t)applicationExtensions.size();
+        computeInstanceCreateInfo.ppEnabledExtensionNames =
+          applicationExtensions.data();
+    }
+
+#ifndef KOMPUTE_DISABLE_VK_DEBUG_LAYERS
+    KP_LOG_DEBUG("Kompute Manager adding debug validation layers");
+    // We'll identify the layers that are supported
+    std::vector<const char*> validLayerNames;
+    std::vector<const char*> desiredLayerNames = {
+        "VK_LAYER_LUNARG_assistant_layer",
+        "VK_LAYER_LUNARG_standard_validation",
+        "VK_LAYER_KHRONOS_validation",
+    };
+    std::vector<std::string> envLayerNames;
+    const char* envLayerNamesVal = std::getenv("KOMPUTE_ENV_DEBUG_LAYERS");
+    if (envLayerNamesVal != nullptr && *envLayerNamesVal != '\0') {
+        KP_LOG_DEBUG("Kompute Manager adding environment layers: {}",
+                     envLayerNamesVal);
+        std::istringstream iss(envLayerNamesVal);
+        std::istream_iterator<std::string> beg(iss);
+        std::istream_iterator<std::string> end;
+        envLayerNames = std::vector<std::string>(beg, end);
+        for (const std::string& layerName : envLayerNames) {
+            desiredLayerNames.push_back(layerName.c_str());
+        }
+        KP_LOG_DEBUG("Desired layers: {}", fmt::join(desiredLayerNames, ", "));
+    }
+
+    // Identify the valid layer names based on the desiredLayerNames
+    {
+        std::set<std::string> uniqueLayerNames;
+        std::vector<vk::LayerProperties> availableLayerProperties =
+          vk::enumerateInstanceLayerProperties();
+        for (vk::LayerProperties layerProperties : availableLayerProperties) {
+            std::string layerName(layerProperties.layerName.data());
+            uniqueLayerNames.insert(layerName);
+        }
+        KP_LOG_DEBUG("Available layers: {}", fmt::join(uniqueLayerNames, ", "));
+        for (const char* desiredLayerName : desiredLayerNames) {
+            if (uniqueLayerNames.count(desiredLayerName) != 0) {
+                validLayerNames.push_back(desiredLayerName);
+            }
+        }
+    }
+
+    if (!validLayerNames.empty()) {
+        KP_LOG_DEBUG(
+          "Kompute Manager Initializing instance with valid layers: {}",
+          fmt::join(validLayerNames, ", "));
+        computeInstanceCreateInfo.enabledLayerCount =
+          static_cast<uint32_t>(validLayerNames.size());
+        computeInstanceCreateInfo.ppEnabledLayerNames = validLayerNames.data();
+    } else {
+        KP_LOG_WARN("Kompute Manager no valid layer names found from desired "
+                    "layer names");
+    }
+#endif
+
+    try {
+        mDynamicLoader = std::make_shared<vk::DynamicLoader>();
+    } catch (const std::exception & err) {
+        return;
+    }
+
+    PFN_vkGetInstanceProcAddr vkGetInstanceProcAddr =
+      mDynamicLoader->getProcAddress<PFN_vkGetInstanceProcAddr>("vkGetInstanceProcAddr");
+    VULKAN_HPP_DEFAULT_DISPATCHER.init(vkGetInstanceProcAddr);
+
+    this->mInstance = std::make_shared<vk::Instance>();
+    vk::createInstance(
+      &computeInstanceCreateInfo, nullptr, this->mInstance.get());
+
+    VULKAN_HPP_DEFAULT_DISPATCHER.init(*this->mInstance);
+
+    KP_LOG_DEBUG("Kompute Manager Instance Created");
+
+#ifndef KOMPUTE_DISABLE_VK_DEBUG_LAYERS
+    KP_LOG_DEBUG("Kompute Manager adding debug callbacks");
+    if (validLayerNames.size() > 0) {
+        vk::DebugReportFlagsEXT debugFlags =
+          vk::DebugReportFlagBitsEXT::eError |
+          vk::DebugReportFlagBitsEXT::eWarning;
+        vk::DebugReportCallbackCreateInfoEXT debugCreateInfo = {};
+        debugCreateInfo.pfnCallback =
+          (PFN_vkDebugReportCallbackEXT)debugMessageCallback;
+        debugCreateInfo.flags = debugFlags;
+
+        this->mDebugDispatcher.init(*this->mInstance, &vkGetInstanceProcAddr);
+        this->mDebugReportCallback =
+          this->mInstance->createDebugReportCallbackEXT(
+            debugCreateInfo, nullptr, this->mDebugDispatcher);
+    }
+#endif
+}
+
+void
+Manager::clear()
+{
+    if (this->mManageResources) {
+        this->mManagedTensors.erase(
+          std::remove_if(begin(this->mManagedTensors),
+                         end(this->mManagedTensors),
+                         [](std::weak_ptr<Tensor> t) { return t.expired(); }),
+          end(this->mManagedTensors));
+        for (auto it = this->mManagedAlgorithmsMap.begin();
+             it != this->mManagedAlgorithmsMap.end();) {
+            if (it->second) {
+                it = this->mManagedAlgorithmsMap.erase(it);
+            } else {
+                ++it;
+            }
+        }
+        this->mManagedSequences.erase(
+          std::remove_if(begin(this->mManagedSequences),
+                         end(this->mManagedSequences),
+                         [](std::weak_ptr<Sequence> t) { return t.expired(); }),
+          end(this->mManagedSequences));
+    }
+}
+
+void
+Manager::createDevice(const std::vector<uint32_t>& familyQueueIndices,
+                      uint32_t physicalDeviceIndex,
+                      const std::vector<std::string>& desiredExtensions)
+{
+
+    KP_LOG_DEBUG("Kompute Manager creating Device");
+
+    if (this->mInstance == nullptr) {
+        throw std::runtime_error("Kompute Manager instance is null");
+    }
+
+    this->mFreeDevice = true;
+
+    // Getting an integer that says how many vuklan devices we have
+    std::vector<vk::PhysicalDevice> physicalDevices =
+      this->mInstance->enumeratePhysicalDevices();
+    uint32_t deviceCount = physicalDevices.size();
+
+    // This means there are no devices at all
+    if (deviceCount == 0) {
+        throw std::runtime_error("Failed to find GPUs with Vulkan support! "
+                                 "Maybe you haven't installed vulkan drivers?");
+    }
+
+    // This means that we're exceeding our device limit, for
+    // example if we have 2 devices, just physicalDeviceIndex
+    // 0 and 1 are acceptable. Hence, physicalDeviceIndex should
+    // always be less than deviceCount, else we raise an error
+    if (!(deviceCount > physicalDeviceIndex)) {
+        throw std::runtime_error("There is no such physical index or device, "
+                                 "please use your existing device");
+    }
+
+    vk::PhysicalDevice physicalDevice = physicalDevices[physicalDeviceIndex];
+
+    this->mPhysicalDevice =
+      std::make_shared<vk::PhysicalDevice>(physicalDevice);
+
+#if KOMPUTE_OPT_ACTIVE_LOG_LEVEL <= KOMPUTE_LOG_LEVEL_INFO
+    vk::PhysicalDeviceProperties physicalDeviceProperties =
+      physicalDevice.getProperties();
+#endif
+
+    KP_LOG_INFO("Using physical device index {} found {}",
+                physicalDeviceIndex,
+                physicalDeviceProperties.deviceName);
+
+    if (familyQueueIndices.empty()) {
+        // Find compute queue
+        std::vector<vk::QueueFamilyProperties> allQueueFamilyProperties =
+          physicalDevice.getQueueFamilyProperties();
+
+        uint32_t computeQueueFamilyIndex = 0;
+        bool computeQueueSupported = false;
+        for (uint32_t i = 0; i < allQueueFamilyProperties.size(); i++) {
+            vk::QueueFamilyProperties queueFamilyProperties =
+              allQueueFamilyProperties[i];
+
+            if (queueFamilyProperties.queueFlags &
+                vk::QueueFlagBits::eCompute) {
+                computeQueueFamilyIndex = i;
+                computeQueueSupported = true;
+                break;
+            }
+        }
+
+        if (!computeQueueSupported) {
+            throw std::runtime_error("Compute queue is not supported");
+        }
+
+        this->mComputeQueueFamilyIndices.push_back(computeQueueFamilyIndex);
+    } else {
+        this->mComputeQueueFamilyIndices = familyQueueIndices;
+    }
+
+    std::unordered_map<uint32_t, uint32_t> familyQueueCounts;
+    std::unordered_map<uint32_t, std::vector<float>> familyQueuePriorities;
+    for (const auto& value : this->mComputeQueueFamilyIndices) {
+        familyQueueCounts[value]++;
+        familyQueuePriorities[value].push_back(1.0f);
+    }
+
+    std::unordered_map<uint32_t, uint32_t> familyQueueIndexCount;
+    std::vector<vk::DeviceQueueCreateInfo> deviceQueueCreateInfos;
+    for (const auto& familyQueueInfo : familyQueueCounts) {
+        // Setting the device count to 0
+        familyQueueIndexCount[familyQueueInfo.first] = 0;
+
+        // Creating the respective device queue
+        vk::DeviceQueueCreateInfo deviceQueueCreateInfo(
+          vk::DeviceQueueCreateFlags(),
+          familyQueueInfo.first,
+          familyQueueInfo.second,
+          familyQueuePriorities[familyQueueInfo.first].data());
+        deviceQueueCreateInfos.push_back(deviceQueueCreateInfo);
+    }
+
+    KP_LOG_DEBUG("Kompute Manager desired extension layers {}",
+                 fmt::join(desiredExtensions, ", "));
+
+    std::vector<vk::ExtensionProperties> deviceExtensions =
+      this->mPhysicalDevice->enumerateDeviceExtensionProperties();
+
+    std::set<std::string> uniqueExtensionNames;
+    for (const vk::ExtensionProperties& ext : deviceExtensions) {
+        uniqueExtensionNames.insert(ext.extensionName);
+    }
+    KP_LOG_DEBUG("Kompute Manager available extensions {}",
+                 fmt::join(uniqueExtensionNames, ", "));
+    std::vector<const char*> validExtensions;
+    for (const std::string& ext : desiredExtensions) {
+        if (uniqueExtensionNames.count(ext) != 0) {
+            validExtensions.push_back(ext.c_str());
+        }
+    }
+    if (desiredExtensions.size() != validExtensions.size()) {
+        KP_LOG_ERROR("Kompute Manager not all extensions were added: {}",
+                     fmt::join(validExtensions, ", "));
+    }
+
+    vk::PhysicalDeviceFeatures features;
+    features.shaderInt16 = true;
+
+    vk::PhysicalDeviceVulkan11Features features11;
+    features11.uniformAndStorageBuffer16BitAccess = true;
+    features11.storageBuffer16BitAccess = true;
+    features11.pNext = nullptr;
+
+    vk::PhysicalDeviceVulkan12Features features12;
+    features12.storageBuffer8BitAccess = true;
+    features12.uniformAndStorageBuffer8BitAccess = true;
+    features12.shaderFloat16 = true;
+    features12.shaderInt8 = true;
+    features12.pNext = &features11;
+
+    vk::DeviceCreateInfo deviceCreateInfo(vk::DeviceCreateFlags(),
+                                          deviceQueueCreateInfos.size(),
+                                          deviceQueueCreateInfos.data(),
+                                          {},
+                                          {},
+                                          validExtensions.size(),
+                                          validExtensions.data(),
+                                          &features);
+
+    deviceCreateInfo.pNext = &features12;
+
+    this->mDevice = std::make_shared<vk::Device>();
+    vk::Result r = physicalDevice.createDevice(
+      &deviceCreateInfo, nullptr, this->mDevice.get());
+    if (r != vk::Result::eSuccess) {
+        KP_LOG_ERROR("Kompute Manager could not create device");
+    }
+
+    KP_LOG_DEBUG("Kompute Manager device created");
+
+    for (const uint32_t& familyQueueIndex : this->mComputeQueueFamilyIndices) {
+        std::shared_ptr<vk::Queue> currQueue = std::make_shared<vk::Queue>();
+
+        this->mDevice->getQueue(familyQueueIndex,
+                                familyQueueIndexCount[familyQueueIndex],
+                                currQueue.get());
+
+        familyQueueIndexCount[familyQueueIndex]++;
+
+        this->mComputeQueues.push_back(currQueue);
+    }
+
+    KP_LOG_DEBUG("Kompute Manager compute queue obtained");
+
+    mPipelineCache = std::make_shared<vk::PipelineCache>();
+    vk::PipelineCacheCreateInfo pipelineCacheInfo =
+        vk::PipelineCacheCreateInfo();
+    this->mDevice->createPipelineCache(
+        &pipelineCacheInfo, nullptr, mPipelineCache.get());
+}
+
+std::shared_ptr<Sequence>
+Manager::sequence(uint32_t queueIndex, uint32_t totalTimestamps)
+{
+    KP_LOG_DEBUG("Kompute Manager sequence() with queueIndex: {}", queueIndex);
+
+    std::shared_ptr<Sequence> sq{ new kp::Sequence(
+      this->mPhysicalDevice,
+      this->mDevice,
+      this->mComputeQueues[queueIndex],
+      this->mComputeQueueFamilyIndices[queueIndex],
+      totalTimestamps) };
+
+    if (this->mManageResources) {
+        this->mManagedSequences.push_back(sq);
+    }
+
+    return sq;
+}
+
+vk::PhysicalDeviceProperties
+Manager::getDeviceProperties() const
+{
+    return this->mPhysicalDevice->getProperties();
+}
+
+std::vector<vk::PhysicalDevice>
+Manager::listDevices() const
+{
+    return this->mInstance->enumeratePhysicalDevices();
+}
+
+std::shared_ptr<vk::Instance>
+Manager::getVkInstance() const
+{
+    return this->mInstance;
+}
+
+}

kompute/src/OpAlgoDispatch.cpp

@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/operations/OpAlgoDispatch.hpp"
+
+namespace kp {
+
+OpAlgoDispatch::~OpAlgoDispatch()
+{
+    KP_LOG_DEBUG("Kompute OpAlgoDispatch destructor started");
+
+    if (this->mPushConstantsData) {
+        KP_LOG_DEBUG("Kompute freeing push constants data");
+        free(this->mPushConstantsData);
+    }
+}
+
+void
+OpAlgoDispatch::record(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute OpAlgoDispatch record called");
+
+    // Barrier to ensure the data is finished writing to buffer memory
+    for (const std::shared_ptr<Tensor>& tensor :
+         this->mAlgorithm->getTensors()) {
+        tensor->recordPrimaryBufferMemoryBarrier(
+          commandBuffer,
+          vk::AccessFlagBits::eTransferWrite,
+          vk::AccessFlagBits::eShaderRead,
+          vk::PipelineStageFlagBits::eTransfer,
+          vk::PipelineStageFlagBits::eComputeShader);
+    }
+
+    if (this->mPushConstantsSize) {
+        this->mAlgorithm->setPushConstants(
+          this->mPushConstantsData,
+          this->mPushConstantsSize,
+          this->mPushConstantsDataTypeMemorySize);
+    }
+
+    this->mAlgorithm->recordBindCore(commandBuffer);
+    this->mAlgorithm->recordBindPush(commandBuffer);
+    this->mAlgorithm->recordDispatch(commandBuffer);
+}
+
+void
+OpAlgoDispatch::preEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpAlgoDispatch preEval called");
+}
+
+void
+OpAlgoDispatch::postEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpAlgoDispatch postSubmit called");
+}
+
+}

kompute/src/OpBufferSyncDevice.cpp

@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/operations/OpBufferSyncDevice.hpp"
+
+namespace kp {
+
+OpBufferSyncDevice::OpBufferSyncDevice(
+        vk::Buffer *primaryBuffer,
+        vk::Buffer *stagingBuffer,
+        vk::DeviceSize size)
+  : mPrimaryBuffer(primaryBuffer)
+  , mStagingBuffer(stagingBuffer)
+  , mSize(size)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncDevice constructor with params");
+}
+
+OpBufferSyncDevice::~OpBufferSyncDevice()
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncDevice destructor started");
+}
+
+void
+OpBufferSyncDevice::record(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncDevice record called");
+    vk::BufferCopy copyRegion(0, 0, mSize);
+    commandBuffer.copyBuffer(*mStagingBuffer, *mPrimaryBuffer, copyRegion);
+}
+
+void
+OpBufferSyncDevice::preEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncDevice preEval called");
+}
+
+void
+OpBufferSyncDevice::postEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncDevice postEval called");
+}
+
+}

kompute/src/OpBufferSyncLocal.cpp

@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/operations/OpBufferSyncLocal.hpp"
+
+namespace kp {
+
+OpBufferSyncLocal::OpBufferSyncLocal(
+        vk::Buffer *primaryBuffer,
+        vk::Buffer *stagingBuffer,
+        vk::DeviceSize size)
+  : mPrimaryBuffer(primaryBuffer)
+  , mStagingBuffer(stagingBuffer)
+  , mSize(size)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncLocal constructor with params");
+}
+
+OpBufferSyncLocal::~OpBufferSyncLocal()
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncLocal destructor started");
+}
+
+void
+OpBufferSyncLocal::record(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncLocal record called");
+    vk::BufferCopy copyRegion(0, 0, mSize);
+    commandBuffer.copyBuffer(*mPrimaryBuffer, *mStagingBuffer, copyRegion);
+}
+
+void
+OpBufferSyncLocal::preEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncLocal preEval called");
+}
+
+void
+OpBufferSyncLocal::postEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpBufferSyncLocal postEval called");
+}
+
+}

kompute/src/OpMemoryBarrier.cpp

@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/operations/OpMemoryBarrier.hpp"
+
+namespace kp {
+
+OpMemoryBarrier::OpMemoryBarrier(
+  const std::vector<std::shared_ptr<Tensor>>& tensors,
+  const vk::AccessFlagBits& srcAccessMask,
+  const vk::AccessFlagBits& dstAccessMask,
+  const vk::PipelineStageFlagBits& srcStageMask,
+  const vk::PipelineStageFlagBits& dstStageMask,
+  bool barrierOnPrimary)
+  : mSrcAccessMask(srcAccessMask)
+  , mDstAccessMask(dstAccessMask)
+  , mSrcStageMask(srcStageMask)
+  , mDstStageMask(dstStageMask)
+  , mBarrierOnPrimary(barrierOnPrimary)
+  , mTensors(tensors)
+{
+    KP_LOG_DEBUG("Kompute OpMemoryBarrier constructor");
+}
+
+OpMemoryBarrier::~OpMemoryBarrier()
+{
+    KP_LOG_DEBUG("Kompute OpMemoryBarrier destructor started");
+}
+
+void
+OpMemoryBarrier::record(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute OpMemoryBarrier record called");
+
+    // Barrier to ensure the data is finished writing to buffer memory
+    if (this->mBarrierOnPrimary) {
+        for (const std::shared_ptr<Tensor>& tensor : this->mTensors) {
+            tensor->recordPrimaryBufferMemoryBarrier(commandBuffer,
+                                                     this->mSrcAccessMask,
+                                                     this->mDstAccessMask,
+                                                     this->mSrcStageMask,
+                                                     this->mDstStageMask);
+        }
+    } else {
+        for (const std::shared_ptr<Tensor>& tensor : this->mTensors) {
+            tensor->recordStagingBufferMemoryBarrier(commandBuffer,
+                                                     this->mSrcAccessMask,
+                                                     this->mDstAccessMask,
+                                                     this->mSrcStageMask,
+                                                     this->mDstStageMask);
+        }
+    }
+}
+
+void
+OpMemoryBarrier::preEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpMemoryBarrier preEval called");
+}
+
+void
+OpMemoryBarrier::postEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpMemoryBarrier postSubmit called");
+}
+
+}

kompute/src/OpTensorCopy.cpp

@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/operations/OpTensorCopy.hpp"
+#include "kompute/Tensor.hpp"
+
+namespace kp {
+
+OpTensorCopy::OpTensorCopy(const std::vector<std::shared_ptr<Tensor>>& tensors)
+{
+    KP_LOG_DEBUG("Kompute OpTensorCopy constructor with params");
+
+    this->mTensors = tensors;
+
+    if (this->mTensors.size() < 2) {
+        throw std::runtime_error(
+          "Kompute OpTensorCopy called with less than 2 tensor");
+    }
+
+    kp::Tensor::TensorDataTypes dataType = this->mTensors[0]->dataType();
+    uint32_t size = this->mTensors[0]->size();
+    for (const std::shared_ptr<Tensor>& tensor : tensors) {
+        if (tensor->dataType() != dataType) {
+            throw std::runtime_error(fmt::format(
+              "Attempting to copy tensors of different types from {} to {}",
+              Tensor::toString(dataType),
+              Tensor::toString(tensor->dataType())));
+        }
+        if (tensor->size() != size) {
+            throw std::runtime_error(fmt::format(
+              "Attempting to copy tensors of different sizes from {} to {}",
+              size,
+              tensor->size()));
+        }
+    }
+}
+
+OpTensorCopy::~OpTensorCopy()
+{
+    KP_LOG_DEBUG("Kompute OpTensorCopy destructor started");
+}
+
+void
+OpTensorCopy::record(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute OpTensorCopy record called");
+
+    // We iterate from the second tensor onwards and record a copy to all
+    for (size_t i = 1; i < this->mTensors.size(); i++) {
+        this->mTensors[i]->recordCopyFrom(commandBuffer, this->mTensors[0]);
+    }
+}
+
+void
+OpTensorCopy::preEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpTensorCopy preEval called");
+}
+
+void
+OpTensorCopy::postEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpTensorCopy postEval called");
+
+    // Do not copy on CPU side if source is storage tensor
+    if (this->mTensors[0]->tensorType() == kp::Tensor::TensorTypes::eStorage)
+    {
+        KP_LOG_DEBUG("Kompute OpTensorCopy not copying tensor source given it's of eStorage type");
+        return;
+    }
+    void* data = this->mTensors[0]->rawData();
+
+    // Copy the data from the first tensor into all the tensors
+    for (size_t i = 1; i < this->mTensors.size(); i++) {
+        if (this->mTensors[i]->tensorType() == kp::Tensor::TensorTypes::eStorage) {
+            KP_LOG_DEBUG("Kompute OpTensorCopy not copying to tensor dest given it's of eStorage type");
+            continue;
+        }
+        this->mTensors[i]->setRawData(data);
+    }
+}
+
+}

kompute/src/OpTensorSyncDevice.cpp

@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/operations/OpTensorSyncDevice.hpp"
+
+namespace kp {
+
+OpTensorSyncDevice::OpTensorSyncDevice(
+  const std::vector<std::shared_ptr<Tensor>>& tensors)
+  : mPrimaryBuffer(nullptr)
+  , mStagingBuffer(nullptr)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncDevice constructor with params");
+
+    if (tensors.size() < 1) {
+        throw std::runtime_error(
+          "Kompute OpTensorSyncDevice called with less than 1 tensor");
+    }
+
+    this->mTensors = tensors;
+}
+
+OpTensorSyncDevice::~OpTensorSyncDevice()
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncDevice destructor started");
+
+    this->mTensors.clear();
+}
+
+void
+OpTensorSyncDevice::record(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncDevice record called");
+
+    for (size_t i = 0; i < this->mTensors.size(); i++) {
+        if (this->mTensors[i]->tensorType() == Tensor::TensorTypes::eDevice) {
+            this->mTensors[i]->recordCopyFromStagingToDevice(commandBuffer);
+        }
+    }
+}
+
+void
+OpTensorSyncDevice::preEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncDevice preEval called");
+}
+
+void
+OpTensorSyncDevice::postEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncDevice postEval called");
+}
+
+}

kompute/src/OpTensorSyncLocal.cpp

@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Copyright (c) 2023 Nomic, Inc. All rights reserved.
+ *
+ * This software is licensed under the terms of the Software for Open Models License (SOM),
+ * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
+ * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
+ */
+
+#include "kompute/Tensor.hpp"
+
+#include "kompute/operations/OpTensorSyncLocal.hpp"
+
+namespace kp {
+
+OpTensorSyncLocal::OpTensorSyncLocal(
+  const std::vector<std::shared_ptr<Tensor>>& tensors)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncLocal constructor with params");
+
+    if (tensors.size() < 1) {
+        throw std::runtime_error(
+          "Kompute OpTensorSyncLocal called with less than 1 tensor");
+    }
+
+    this->mTensors = tensors;
+}
+
+OpTensorSyncLocal::~OpTensorSyncLocal()
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncLocal destructor started");
+}
+
+void
+OpTensorSyncLocal::record(const vk::CommandBuffer& commandBuffer)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncLocal record called");
+
+    for (size_t i = 0; i < this->mTensors.size(); i++) {
+        if (this->mTensors[i]->tensorType() == Tensor::TensorTypes::eDevice) {
+
+            this->mTensors[i]->recordPrimaryBufferMemoryBarrier(
+              commandBuffer,
+              vk::AccessFlagBits::eShaderWrite,
+              vk::AccessFlagBits::eTransferRead,
+              vk::PipelineStageFlagBits::eComputeShader,
+              vk::PipelineStageFlagBits::eTransfer);
+
+            this->mTensors[i]->recordCopyFromDeviceToStaging(commandBuffer);
+
+            this->mTensors[i]->recordPrimaryBufferMemoryBarrier(
+              commandBuffer,
+              vk::AccessFlagBits::eTransferWrite,
+              vk::AccessFlagBits::eHostRead,
+              vk::PipelineStageFlagBits::eTransfer,
+              vk::PipelineStageFlagBits::eHost);
+        }
+    }
+}
+
+void
+OpTensorSyncLocal::preEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncLocal preEval called");
+}
+
+void
+OpTensorSyncLocal::postEval(const vk::CommandBuffer& /*commandBuffer*/)
+{
+    KP_LOG_DEBUG("Kompute OpTensorSyncLocal postEval called");
+
+    KP_LOG_DEBUG("Kompute OpTensorSyncLocal mapping data into tensor local");
+}
+
+}